Jon Gauthier
commited on
Commit
·
452d201
1
Parent(s):
e019438
extract metric code for SyntaxGym metric space
Browse files- prediction.py +0 -235
- syntaxgym.py +0 -184
- test.py +0 -515
prediction.py
DELETED
@@ -1,235 +0,0 @@
|
|
1 |
-
from typing import Union, Optional as TOptional, List as TList
|
2 |
-
|
3 |
-
|
4 |
-
from pyparsing import *
|
5 |
-
import numpy as np
|
6 |
-
|
7 |
-
METRICS = {
|
8 |
-
'sum': sum,
|
9 |
-
'mean': np.mean,
|
10 |
-
'median': np.median,
|
11 |
-
'range': np.ptp,
|
12 |
-
'max': max,
|
13 |
-
'min': min
|
14 |
-
}
|
15 |
-
|
16 |
-
|
17 |
-
# Enable parser packrat (caching)
|
18 |
-
ParserElement.enablePackrat()
|
19 |
-
|
20 |
-
# Relative and absolute tolerance thresholds for surprisal equality
|
21 |
-
EQUALITY_RTOL = 1e-5
|
22 |
-
EQUALITY_ATOL = 1e-3
|
23 |
-
|
24 |
-
|
25 |
-
#######
|
26 |
-
# Define a grammar for prediction formulae.
|
27 |
-
|
28 |
-
# References a surprisal region
|
29 |
-
lpar = Suppress("(")
|
30 |
-
rpar = Suppress(")")
|
31 |
-
region = lpar + (Word(nums) | "*") + Suppress(";%") + Word(alphanums + "_-") + Suppress("%") + rpar
|
32 |
-
literal_float = pyparsing_common.number
|
33 |
-
|
34 |
-
class Region(object):
|
35 |
-
def __init__(self, tokens):
|
36 |
-
self.region_number = tokens[0]
|
37 |
-
self.condition_name = tokens[1]
|
38 |
-
|
39 |
-
def __str__(self):
|
40 |
-
return "(%s;%%%s%%)" % (self.region_number, self.condition_name)
|
41 |
-
|
42 |
-
def __repr__(self):
|
43 |
-
return "Region(%s,%s)" % (self.condition_name, self.region_number)
|
44 |
-
|
45 |
-
def __call__(self, surprisal_dict):
|
46 |
-
if self.region_number == "*":
|
47 |
-
return sum(value for (condition, region), value in surprisal_dict.items()
|
48 |
-
if condition == self.condition_name)
|
49 |
-
|
50 |
-
return surprisal_dict[self.condition_name, int(self.region_number)]
|
51 |
-
|
52 |
-
class LiteralFloat(object):
|
53 |
-
def __init__(self, tokens):
|
54 |
-
self.value = float(tokens[0])
|
55 |
-
|
56 |
-
def __str__(self):
|
57 |
-
return "%f" % (self.value,)
|
58 |
-
|
59 |
-
def __repr__(self):
|
60 |
-
return "LiteralFloat(%f)" % (self.value,)
|
61 |
-
|
62 |
-
def __call__(self, surprisal_dict):
|
63 |
-
return self.value
|
64 |
-
|
65 |
-
class BinaryOp(object):
|
66 |
-
operators: TOptional[TList[str]]
|
67 |
-
|
68 |
-
def __init__(self, tokens):
|
69 |
-
self.operator = tokens[0][1]
|
70 |
-
if self.operators is not None and self.operator not in self.operators:
|
71 |
-
raise ValueError("Invalid %s operator %s" % (self.__class__.__name__,
|
72 |
-
self.operator))
|
73 |
-
self.operands = [tokens[0][0], tokens[0][2]]
|
74 |
-
|
75 |
-
def __str__(self):
|
76 |
-
return "(%s %s %s)" % (self.operands[0], self.operator, self.operands[1])
|
77 |
-
|
78 |
-
def __repr__(self):
|
79 |
-
return "%s(%s)(%s)" % (self.__class__.__name__, self.operator, ",".join(map(repr, self.operands)))
|
80 |
-
|
81 |
-
def __call__(self, surprisal_dict):
|
82 |
-
op_vals = [op(surprisal_dict) for op in self.operands]
|
83 |
-
return self._evaluate(op_vals, surprisal_dict)
|
84 |
-
|
85 |
-
def _evaluate(self, evaluated_operands, surprisal_dict):
|
86 |
-
raise NotImplementedError()
|
87 |
-
|
88 |
-
class BoolOp(BinaryOp):
|
89 |
-
operators = ["&", "|"]
|
90 |
-
def _evaluate(self, op_vals, surprisal_dict):
|
91 |
-
if self.operator == "&":
|
92 |
-
return op_vals[0] and op_vals[1]
|
93 |
-
elif self.operator == "|":
|
94 |
-
return op_vals[0] or op_vals[1]
|
95 |
-
|
96 |
-
class FloatOp(BinaryOp):
|
97 |
-
operators = ["-", "+"]
|
98 |
-
def _evaluate(self, op_vals, surprisal_dict):
|
99 |
-
if self.operator == "-":
|
100 |
-
return op_vals[0] - op_vals[1]
|
101 |
-
elif self.operator == "+":
|
102 |
-
return op_vals[0] + op_vals[1]
|
103 |
-
|
104 |
-
class ComparatorOp(BinaryOp):
|
105 |
-
operators = ["<", ">", "="]
|
106 |
-
def _evaluate(self, op_vals, surprisal_dict):
|
107 |
-
if self.operator == "<":
|
108 |
-
return op_vals[0] < op_vals[1]
|
109 |
-
elif self.operator == ">":
|
110 |
-
return op_vals[0] > op_vals[1]
|
111 |
-
elif self.operator == "=":
|
112 |
-
return np.isclose(op_vals[0], op_vals[1],
|
113 |
-
rtol=EQUALITY_RTOL,
|
114 |
-
atol=EQUALITY_ATOL)
|
115 |
-
|
116 |
-
def Chain(op_cls, left_assoc=True):
|
117 |
-
def chainer(tokens):
|
118 |
-
"""
|
119 |
-
Create a binary tree of BinaryOps from the given repeated application
|
120 |
-
of the op.
|
121 |
-
"""
|
122 |
-
operators = tokens[0][1::2]
|
123 |
-
args = tokens[0][0::2]
|
124 |
-
if not left_assoc:
|
125 |
-
raise NotImplementedError
|
126 |
-
|
127 |
-
arg1 = args.pop(0)
|
128 |
-
while len(args) > 0:
|
129 |
-
operator = operators.pop(0)
|
130 |
-
arg2 = args.pop(0)
|
131 |
-
arg1 = op_cls([[arg1, operator, arg2]])
|
132 |
-
|
133 |
-
return arg1
|
134 |
-
|
135 |
-
return chainer
|
136 |
-
|
137 |
-
atom = region.setParseAction(Region) | literal_float.setParseAction(LiteralFloat)
|
138 |
-
|
139 |
-
prediction_expr = infixNotation(
|
140 |
-
atom,
|
141 |
-
[
|
142 |
-
(oneOf("- +"), 2, opAssoc.LEFT, Chain(FloatOp)),
|
143 |
-
(oneOf("< > ="), 2, opAssoc.LEFT, ComparatorOp),
|
144 |
-
(oneOf("& |"), 2, opAssoc.LEFT, Chain(BoolOp)),
|
145 |
-
],
|
146 |
-
lpar=lpar, rpar=rpar
|
147 |
-
)
|
148 |
-
|
149 |
-
|
150 |
-
class Prediction(object):
|
151 |
-
"""
|
152 |
-
Predictions state expected relations between language model surprisal
|
153 |
-
measures in different regions and conditions of a test suite. For more
|
154 |
-
information, see :ref:`architecture`.
|
155 |
-
"""
|
156 |
-
|
157 |
-
def __init__(self, idx: int, formula: Union[str, BinaryOp], metric: str):
|
158 |
-
"""
|
159 |
-
Args:
|
160 |
-
idx: A unique prediction ID. This is only relevant for
|
161 |
-
serialization.
|
162 |
-
formula: A string representation of the prediction formula, or an
|
163 |
-
already parsed formula. For more information, see
|
164 |
-
:ref:`architecture`.
|
165 |
-
metric: Metric for aggregating surprisals within regions.
|
166 |
-
"""
|
167 |
-
if isinstance(formula, str):
|
168 |
-
try:
|
169 |
-
formula = prediction_expr.parseString(formula, parseAll=True)[0]
|
170 |
-
except ParseException as e:
|
171 |
-
raise ValueError("Invalid formula expression %r" % (formula,)) from e
|
172 |
-
|
173 |
-
self.idx = idx
|
174 |
-
self.formula = formula
|
175 |
-
|
176 |
-
if metric not in METRICS.keys():
|
177 |
-
raise ValueError("Unknown metric %s. Supported metrics: %s" %
|
178 |
-
(metric, " ".join(METRICS.keys())))
|
179 |
-
self.metric = metric
|
180 |
-
|
181 |
-
def __call__(self, item):
|
182 |
-
"""
|
183 |
-
Evaluate the prediction on the given item dict representation. For more
|
184 |
-
information on item representations, see :ref:`suite_json`.
|
185 |
-
"""
|
186 |
-
# Prepare relevant surprisal dict
|
187 |
-
surps = {(c["condition_name"], r["region_number"]): r["metric_value"][self.metric]
|
188 |
-
for c in item["conditions"]
|
189 |
-
for r in c["regions"]}
|
190 |
-
return self.formula(surps)
|
191 |
-
|
192 |
-
@classmethod
|
193 |
-
def from_dict(cls, pred_dict, idx: int, metric: str):
|
194 |
-
"""
|
195 |
-
Parse from a prediction dictionary representation (see
|
196 |
-
:ref:`suite_json`).
|
197 |
-
"""
|
198 |
-
if not pred_dict["type"] == "formula":
|
199 |
-
raise ValueError("Unknown prediction type %s" % (pred_dict["type"],))
|
200 |
-
|
201 |
-
return cls(formula=pred_dict["formula"], idx=idx, metric=metric)
|
202 |
-
|
203 |
-
@property
|
204 |
-
def referenced_regions(self):
|
205 |
-
"""
|
206 |
-
Get a set of the regions referenced by this formula.
|
207 |
-
Each item is a tuple of the form ``(condition_name, region_number)``.
|
208 |
-
"""
|
209 |
-
def traverse(x, acc):
|
210 |
-
if isinstance(x, BinaryOp):
|
211 |
-
for val in x.operands:
|
212 |
-
traverse(val, acc)
|
213 |
-
elif isinstance(x, Region):
|
214 |
-
acc.add((x.condition_name, int(x.region_number)))
|
215 |
-
|
216 |
-
return acc
|
217 |
-
|
218 |
-
return traverse(self.formula, set())
|
219 |
-
|
220 |
-
def as_dict(self):
|
221 |
-
"""
|
222 |
-
Serialize as a prediction dictionary representation (see
|
223 |
-
:ref:`suite_json`).
|
224 |
-
"""
|
225 |
-
return dict(type="formula", formula=str(self.formula))
|
226 |
-
|
227 |
-
def __str__(self):
|
228 |
-
return "Prediction(%s)" % (self.formula,)
|
229 |
-
__repr__ = __str__
|
230 |
-
|
231 |
-
def __hash__(self):
|
232 |
-
return hash(self.formula)
|
233 |
-
|
234 |
-
def __eq__(self, other):
|
235 |
-
return isinstance(other, Prediction) and hash(self) == hash(other)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
syntaxgym.py
CHANGED
@@ -19,8 +19,6 @@ import numpy as np
|
|
19 |
import torch
|
20 |
from transformers import AutoModelForCausalLM, AutoTokenizer
|
21 |
|
22 |
-
from .prediction import Prediction
|
23 |
-
|
24 |
|
25 |
_CITATION = """
|
26 |
@inproceedings{Hu:et-al:2020,
|
@@ -126,185 +124,3 @@ class SyntaxGym(datasets.GeneratorBasedBuilder):
|
|
126 |
|
127 |
yield item["item_number"], item
|
128 |
|
129 |
-
|
130 |
-
class SyntaxGymMetricResult(TypedDict):
|
131 |
-
prediction_results: List[List[bool]]
|
132 |
-
region_totals: List[Dict[Tuple[str, int], float]]
|
133 |
-
|
134 |
-
|
135 |
-
class SyntaxGymMetric(datasets.Metric):
|
136 |
-
"""
|
137 |
-
SyntaxGym prediction evaluation metric.
|
138 |
-
"""
|
139 |
-
|
140 |
-
def _info(self):
|
141 |
-
seq = datasets.Sequence
|
142 |
-
features = datasets.Features({
|
143 |
-
"suite": SUITE_DATASET_SPEC
|
144 |
-
})
|
145 |
-
return datasets.MetricInfo(
|
146 |
-
description="TODO",
|
147 |
-
citation=_CITATION,
|
148 |
-
inputs_description="TODO",
|
149 |
-
features=features,
|
150 |
-
)
|
151 |
-
|
152 |
-
def _compute(self, suite, model_id, batch_size: int = 16,
|
153 |
-
add_start_token=True, device=None
|
154 |
-
) -> SyntaxGymMetricResult:
|
155 |
-
if device is not None:
|
156 |
-
assert device in ["gpu", "cpu", "cuda"]
|
157 |
-
if device == "gpu":
|
158 |
-
device = "cuda"
|
159 |
-
else:
|
160 |
-
device = "cuda" if torch.cuda.is_available() else "cpu"
|
161 |
-
|
162 |
-
model = AutoModelForCausalLM.from_pretrained(model_id)
|
163 |
-
model = model.to(device)
|
164 |
-
model.eval()
|
165 |
-
|
166 |
-
tokenizer = AutoTokenizer.from_pretrained(model_id)
|
167 |
-
# TODO copy from perplexity metric
|
168 |
-
tokenizer.pad_token = tokenizer.eos_token
|
169 |
-
|
170 |
-
results = {"prediction_results": [], "region_totals": []}
|
171 |
-
# TODO batch all items together
|
172 |
-
for item in logging.tqdm(suite):
|
173 |
-
result_single = self._compute_single(item, tokenizer, model, device)
|
174 |
-
|
175 |
-
for k in ["prediction_results", "region_totals"]:
|
176 |
-
results[k].append(result_single[k])
|
177 |
-
|
178 |
-
return results
|
179 |
-
|
180 |
-
|
181 |
-
def _compute_single(self, item, tokenizer, model, device):
|
182 |
-
tokenized = tokenizer(item["conditions"]["content"],
|
183 |
-
padding=True,
|
184 |
-
return_tensors="pt",
|
185 |
-
return_offsets_mapping=True).to(device)
|
186 |
-
|
187 |
-
# input_ids: B * T
|
188 |
-
input_ids = tokenized["input_ids"]
|
189 |
-
assert input_ids.ndim == 2
|
190 |
-
|
191 |
-
# Compute sentence level surprisals.
|
192 |
-
with torch.no_grad():
|
193 |
-
# Pre-softmax predictive distribution B * T * V
|
194 |
-
logits = model(input_ids).logits
|
195 |
-
surprisals = -logits.log_softmax(dim=2) / np.log(2)
|
196 |
-
|
197 |
-
# surprisals: B * T * V
|
198 |
-
assert surprisals.ndim == 3
|
199 |
-
|
200 |
-
# Get surprisals of expected words.
|
201 |
-
surps_shifted = surprisals[:, :-1, :]
|
202 |
-
expected_ids = input_ids[:, 1:]
|
203 |
-
|
204 |
-
# TODO: check this logic
|
205 |
-
tt = expected_ids.unsqueeze(2)
|
206 |
-
# reindexed surprisals: B * (T - 1)
|
207 |
-
surprisals = torch.gather(surps_shifted, 2, expected_ids.unsqueeze(2)) \
|
208 |
-
.squeeze(2)
|
209 |
-
# This is the original, which works but not with multiple axes in expected_ids
|
210 |
-
# surprisals = surps_shifted[range(surps_shifted.shape[0]), expected_ids]
|
211 |
-
|
212 |
-
# surprisals is now B * (T - 1)
|
213 |
-
|
214 |
-
#### aggregate
|
215 |
-
condition_names = item["conditions"]["condition_name"]
|
216 |
-
region_totals = {condition_name: defaultdict(float)
|
217 |
-
for condition_name in condition_names}
|
218 |
-
region2tokens = self.compute_region_token_mapping(
|
219 |
-
item, input_ids, tokenized["offset_mapping"])
|
220 |
-
|
221 |
-
for i, (i_cond, i_inputs) in enumerate(zip(condition_names, input_ids)):
|
222 |
-
for region_number, region_tokens in region2tokens[i_cond].items():
|
223 |
-
for token in region_tokens:
|
224 |
-
if token == 0:
|
225 |
-
# surprisal not defined. pass.
|
226 |
-
continue
|
227 |
-
elif token <= surprisals.shape[1]:
|
228 |
-
region_totals[i_cond][region_number] += surprisals[i, token - 1]
|
229 |
-
else:
|
230 |
-
# TODO don't think this is an issue, just should clean
|
231 |
-
# up the aggregation output
|
232 |
-
assert token == surprisals.shape[1], \
|
233 |
-
"%s %s" % (token, surprisals.shape[1])
|
234 |
-
|
235 |
-
region_totals = {(condition_name, region_number): float(total)
|
236 |
-
for condition_name, totals in region_totals.items()
|
237 |
-
for region_number, total in totals.items()}
|
238 |
-
|
239 |
-
results = {
|
240 |
-
"prediction_results": [
|
241 |
-
Prediction(i, formula, "sum").formula(region_totals)
|
242 |
-
for i, formula in enumerate(item["predictions"])
|
243 |
-
],
|
244 |
-
|
245 |
-
"region_totals": region_totals
|
246 |
-
}
|
247 |
-
return results
|
248 |
-
|
249 |
-
def get_region_edges(self, item, condition_idx):
|
250 |
-
"""
|
251 |
-
Get left edge of each region as a character index.
|
252 |
-
"""
|
253 |
-
# NB this is coupled with `condition_to_string` logic of course
|
254 |
-
|
255 |
-
regions = item["conditions"]["regions"][condition_idx]
|
256 |
-
|
257 |
-
idx = 0
|
258 |
-
ret = []
|
259 |
-
for r_idx, region_content in enumerate(regions["content"]):
|
260 |
-
ret.append(idx)
|
261 |
-
|
262 |
-
region_size = len(region_content)
|
263 |
-
if region_content.strip() != "" and r_idx != 0 and not region_content.startswith(","):
|
264 |
-
# Add joining space
|
265 |
-
region_size += 1
|
266 |
-
|
267 |
-
idx += region_size
|
268 |
-
|
269 |
-
return ret
|
270 |
-
|
271 |
-
def compute_region_token_mapping(self, item, input_ids: torch.LongTensor,
|
272 |
-
offset_mapping: List[Tuple[int, int]]
|
273 |
-
) -> Dict[str, Dict[int, List[int]]]:
|
274 |
-
# input_ids: B * T
|
275 |
-
# offset_mapping: B * T * 2
|
276 |
-
# assumes batch is sorted according to item's condition_name order
|
277 |
-
|
278 |
-
condition_names = item["conditions"]["condition_name"]
|
279 |
-
region2tokens = {cond: defaultdict(list) for cond in condition_names}
|
280 |
-
|
281 |
-
max_long = torch.iinfo(torch.int64).max
|
282 |
-
|
283 |
-
input_ids = input_ids.detach()
|
284 |
-
for i_cond, (i_tokens, i_offsets) in enumerate(zip(input_ids, offset_mapping)):
|
285 |
-
region_edges = self.get_region_edges(item, i_cond)
|
286 |
-
|
287 |
-
t_cursor, r_cursor = 0, 0
|
288 |
-
while t_cursor < i_tokens.shape[0]:
|
289 |
-
# token = i_tokens[t_cursor]
|
290 |
-
token_char_start, token_char_end = i_offsets[t_cursor]
|
291 |
-
|
292 |
-
if token_char_start == token_char_end == 0:
|
293 |
-
# This is a padding token. Skip.
|
294 |
-
# TODO what about BOS/EOS? some models incorporate them
|
295 |
-
t_cursor += 1
|
296 |
-
continue
|
297 |
-
|
298 |
-
region_start = region_edges[r_cursor]
|
299 |
-
region_end = region_edges[r_cursor + 1] \
|
300 |
-
if r_cursor + 1 < len(region_edges) else max_long
|
301 |
-
|
302 |
-
# NB region boundaries are left edges, hence the >= here.
|
303 |
-
if token_char_start >= region_end:
|
304 |
-
r_cursor += 1
|
305 |
-
continue
|
306 |
-
|
307 |
-
region2tokens[condition_names[i_cond]][r_cursor + 1].append(t_cursor)
|
308 |
-
t_cursor += 1
|
309 |
-
|
310 |
-
return region2tokens
|
|
|
19 |
import torch
|
20 |
from transformers import AutoModelForCausalLM, AutoTokenizer
|
21 |
|
|
|
|
|
22 |
|
23 |
_CITATION = """
|
24 |
@inproceedings{Hu:et-al:2020,
|
|
|
124 |
|
125 |
yield item["item_number"], item
|
126 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
test.py
DELETED
@@ -1,515 +0,0 @@
|
|
1 |
-
from typing import List
|
2 |
-
|
3 |
-
import datasets
|
4 |
-
import numpy as np
|
5 |
-
|
6 |
-
import pytest
|
7 |
-
|
8 |
-
|
9 |
-
@pytest.fixture(scope="session")
|
10 |
-
def syntaxgym_dataset():
|
11 |
-
return datasets.load_dataset("syntaxgym", "subordination_src-src")
|
12 |
-
|
13 |
-
|
14 |
-
@pytest.fixture(scope="session")
|
15 |
-
def syntaxgym_metric():
|
16 |
-
return datasets.load_metric("syntaxgym")
|
17 |
-
|
18 |
-
|
19 |
-
@pytest.fixture(scope="session")
|
20 |
-
def model_ref():
|
21 |
-
# return "hf-internal-testing/tiny-random-gpt_neo"
|
22 |
-
return "gpt2"
|
23 |
-
|
24 |
-
|
25 |
-
# Reference region surprisals computed with syntaxgym-core.
|
26 |
-
# See notebook in https://colab.research.google.com/drive/1qziyPcu65jffizSPi-ZGHKR0x7BaHFMS#scrollTo=RgtnScy6LLKi .
|
27 |
-
GPT2_SUBORDINATION_SRC_REFERENCE = \
|
28 |
-
[{('no-sub_matrix', 1): 13.151199615123803,
|
29 |
-
('no-sub_matrix', 2): 38.503222716703526,
|
30 |
-
('no-sub_matrix', 3): 27.623861034812286,
|
31 |
-
('no-sub_matrix', 4): 48.831672846038224,
|
32 |
-
('no-sub_matrix', 5): 38.08533699286694,
|
33 |
-
('no-sub_no-matrix', 1): 13.151199615123803,
|
34 |
-
('no-sub_no-matrix', 2): 38.503222716703526,
|
35 |
-
('no-sub_no-matrix', 3): 27.623861034812286,
|
36 |
-
('no-sub_no-matrix', 4): 48.831687980511504,
|
37 |
-
('no-sub_no-matrix', 5): 1.8096143510772873,
|
38 |
-
('sub_matrix', 1): 14.905592916748805,
|
39 |
-
('sub_matrix', 2): 39.06304309956175,
|
40 |
-
('sub_matrix', 3): 26.862648365854433,
|
41 |
-
('sub_matrix', 4): 50.56554401687938,
|
42 |
-
('sub_matrix', 5): 26.532245572980194,
|
43 |
-
('sub_no-matrix', 1): 14.905592916748805,
|
44 |
-
('sub_no-matrix', 2): 39.06304309956175,
|
45 |
-
('sub_no-matrix', 3): 26.862648365854433,
|
46 |
-
('sub_no-matrix', 4): 50.56553438585093,
|
47 |
-
('sub_no-matrix', 5): 7.470089829866611},
|
48 |
-
{('no-sub_matrix', 1): 10.116093820255577,
|
49 |
-
('no-sub_matrix', 2): 20.96513246705127,
|
50 |
-
('no-sub_matrix', 3): 20.02959138986416,
|
51 |
-
('no-sub_matrix', 4): 23.779661397107446,
|
52 |
-
('no-sub_matrix', 5): 33.2560281692696,
|
53 |
-
('no-sub_no-matrix', 1): 10.116093820255577,
|
54 |
-
('no-sub_no-matrix', 2): 20.96513246705127,
|
55 |
-
('no-sub_no-matrix', 3): 20.02959138986416,
|
56 |
-
('no-sub_no-matrix', 4): 23.779661397107446,
|
57 |
-
('no-sub_no-matrix', 5): 1.9449125865631063,
|
58 |
-
('sub_matrix', 1): 13.545157521732826,
|
59 |
-
('sub_matrix', 2): 24.96048395897244,
|
60 |
-
('sub_matrix', 3): 18.609464944317324,
|
61 |
-
('sub_matrix', 4): 23.057566440062317,
|
62 |
-
('sub_matrix', 5): 26.424454285669032,
|
63 |
-
('sub_no-matrix', 1): 13.545157521732826,
|
64 |
-
('sub_no-matrix', 2): 24.96048395897244,
|
65 |
-
('sub_no-matrix', 3): 18.609464944317324,
|
66 |
-
('sub_no-matrix', 4): 23.057566440062317,
|
67 |
-
('sub_no-matrix', 5): 2.807467838359704},
|
68 |
-
{('no-sub_matrix', 1): 11.992867568477442,
|
69 |
-
('no-sub_matrix', 2): 45.813114232935774,
|
70 |
-
('no-sub_matrix', 3): 24.57554828372551,
|
71 |
-
('no-sub_matrix', 4): 45.334025774062916,
|
72 |
-
('no-sub_matrix', 5): 26.208189541862073,
|
73 |
-
('no-sub_no-matrix', 1): 11.992867568477442,
|
74 |
-
('no-sub_no-matrix', 2): 45.813114232935774,
|
75 |
-
('no-sub_no-matrix', 3): 24.57554828372551,
|
76 |
-
('no-sub_no-matrix', 4): 45.33402766587207,
|
77 |
-
('no-sub_no-matrix', 5): 1.8284485151385752,
|
78 |
-
('sub_matrix', 1): 14.219887768799735,
|
79 |
-
('sub_matrix', 2): 46.25055434117979,
|
80 |
-
('sub_matrix', 3): 23.054221678472672,
|
81 |
-
('sub_matrix', 4): 47.08503858470256,
|
82 |
-
('sub_matrix', 5): 22.154772321452022,
|
83 |
-
('sub_no-matrix', 1): 14.219887768799735,
|
84 |
-
('sub_no-matrix', 2): 46.25055434117979,
|
85 |
-
('sub_no-matrix', 3): 23.054221678472672,
|
86 |
-
('sub_no-matrix', 4): 47.08503858470256,
|
87 |
-
('sub_no-matrix', 5): 3.0655133594366757},
|
88 |
-
{('no-sub_matrix', 1): 10.55002943802296,
|
89 |
-
('no-sub_matrix', 2): 52.419810137608856,
|
90 |
-
('no-sub_matrix', 3): 23.30710475332303,
|
91 |
-
('no-sub_matrix', 4): 37.957905964008944,
|
92 |
-
('no-sub_matrix', 5): 29.259648135104936,
|
93 |
-
('no-sub_no-matrix', 1): 10.55002943802296,
|
94 |
-
('no-sub_no-matrix', 2): 52.419810137608856,
|
95 |
-
('no-sub_no-matrix', 3): 23.30710475332303,
|
96 |
-
('no-sub_no-matrix', 4): 37.957905964008944,
|
97 |
-
('no-sub_no-matrix', 5): 1.9632913405649093,
|
98 |
-
('sub_matrix', 1): 15.289384584900025,
|
99 |
-
('sub_matrix', 2): 53.93652737134243,
|
100 |
-
('sub_matrix', 3): 19.43915835312633,
|
101 |
-
('sub_matrix', 4): 36.459591551099386,
|
102 |
-
('sub_matrix', 5): 22.185742699245417,
|
103 |
-
('sub_no-matrix', 1): 15.289384584900025,
|
104 |
-
('sub_no-matrix', 2): 53.93652737134243,
|
105 |
-
('sub_no-matrix', 3): 19.43915835312633,
|
106 |
-
('sub_no-matrix', 4): 36.4595598203003,
|
107 |
-
('sub_no-matrix', 5): 5.707732355645454},
|
108 |
-
{('no-sub_matrix', 1): 23.543723213902986,
|
109 |
-
('no-sub_matrix', 2): 31.967972102825854,
|
110 |
-
('no-sub_matrix', 3): 29.159572978411727,
|
111 |
-
('no-sub_matrix', 4): 36.61365345925747,
|
112 |
-
('no-sub_matrix', 5): 44.576591305970545,
|
113 |
-
('no-sub_no-matrix', 1): 23.543723213902986,
|
114 |
-
('no-sub_no-matrix', 2): 31.967972102825854,
|
115 |
-
('no-sub_no-matrix', 3): 29.159572978411727,
|
116 |
-
('no-sub_no-matrix', 4): 36.61365345925747,
|
117 |
-
('no-sub_no-matrix', 5): 3.2813457388593714,
|
118 |
-
('sub_matrix', 1): 27.118410129310597,
|
119 |
-
('sub_matrix', 2): 33.909617362987866,
|
120 |
-
('sub_matrix', 3): 28.791166362258743,
|
121 |
-
('sub_matrix', 4): 37.24960609010374,
|
122 |
-
('sub_matrix', 5): 31.660933798006262,
|
123 |
-
('sub_no-matrix', 1): 27.118410129310597,
|
124 |
-
('sub_no-matrix', 2): 33.909617362987866,
|
125 |
-
('sub_no-matrix', 3): 28.791166362258743,
|
126 |
-
('sub_no-matrix', 4): 37.24960609010374,
|
127 |
-
('sub_no-matrix', 5): 7.3613541428239015},
|
128 |
-
{('no-sub_matrix', 1): 14.22171869610082,
|
129 |
-
('no-sub_matrix', 2): 30.270423022911977,
|
130 |
-
('no-sub_matrix', 3): 25.973276891204705,
|
131 |
-
('no-sub_matrix', 4): 28.43856735947716,
|
132 |
-
('no-sub_matrix', 5): 57.39887418731055,
|
133 |
-
('no-sub_no-matrix', 1): 14.22171869610082,
|
134 |
-
('no-sub_no-matrix', 2): 30.270423022911977,
|
135 |
-
('no-sub_no-matrix', 3): 25.973276891204705,
|
136 |
-
('no-sub_no-matrix', 4): 28.43856735947716,
|
137 |
-
('no-sub_no-matrix', 5): 1.7127059109344136,
|
138 |
-
('sub_matrix', 1): 16.39289784951447,
|
139 |
-
('sub_matrix', 2): 31.5671111565765,
|
140 |
-
('sub_matrix', 3): 24.54307828171008,
|
141 |
-
('sub_matrix', 4): 29.249645624130757,
|
142 |
-
('sub_matrix', 5): 53.59155769093577,
|
143 |
-
('sub_no-matrix', 1): 16.39289784951447,
|
144 |
-
('sub_no-matrix', 2): 31.5671111565765,
|
145 |
-
('sub_no-matrix', 3): 24.54307828171008,
|
146 |
-
('sub_no-matrix', 4): 29.249645624130757,
|
147 |
-
('sub_no-matrix', 5): 7.225276653947023},
|
148 |
-
{('no-sub_matrix', 1): 13.729688714733188,
|
149 |
-
('no-sub_matrix', 2): 36.018118127225165,
|
150 |
-
('no-sub_matrix', 3): 28.232055923783275,
|
151 |
-
('no-sub_matrix', 4): 44.44634394296659,
|
152 |
-
('no-sub_matrix', 5): 38.277975147059344,
|
153 |
-
('no-sub_no-matrix', 1): 13.729688714733188,
|
154 |
-
('no-sub_no-matrix', 2): 36.018118127225165,
|
155 |
-
('no-sub_no-matrix', 3): 28.232055923783275,
|
156 |
-
('no-sub_no-matrix', 4): 44.44634394296659,
|
157 |
-
('no-sub_no-matrix', 5): 3.0318996942908414,
|
158 |
-
('sub_matrix', 1): 16.93528744674245,
|
159 |
-
('sub_matrix', 2): 36.545024814326574,
|
160 |
-
('sub_matrix', 3): 26.279603445823692,
|
161 |
-
('sub_matrix', 4): 46.501226364074995,
|
162 |
-
('sub_matrix', 5): 32.155418057793035,
|
163 |
-
('sub_no-matrix', 1): 16.93528744674245,
|
164 |
-
('sub_no-matrix', 2): 36.545024814326574,
|
165 |
-
('sub_no-matrix', 3): 26.279603445823692,
|
166 |
-
('sub_no-matrix', 4): 46.501226364074995,
|
167 |
-
('sub_no-matrix', 5): 4.4581122618864155},
|
168 |
-
{('no-sub_matrix', 1): 15.598113737151568,
|
169 |
-
('no-sub_matrix', 2): 56.12543415244172,
|
170 |
-
('no-sub_matrix', 3): 29.755667770007285,
|
171 |
-
('no-sub_matrix', 4): 51.689282097269995,
|
172 |
-
('no-sub_matrix', 5): 45.575230324010775,
|
173 |
-
('no-sub_no-matrix', 1): 15.598113737151568,
|
174 |
-
('no-sub_no-matrix', 2): 56.12543415244172,
|
175 |
-
('no-sub_no-matrix', 3): 29.755667770007285,
|
176 |
-
('no-sub_no-matrix', 4): 51.68928424705313,
|
177 |
-
('no-sub_no-matrix', 5): 1.235207173694806,
|
178 |
-
('sub_matrix', 1): 18.909088991066888,
|
179 |
-
('sub_matrix', 2): 57.753410746636746,
|
180 |
-
('sub_matrix', 3): 28.677667873674363,
|
181 |
-
('sub_matrix', 4): 51.99410775929489,
|
182 |
-
('sub_matrix', 5): 35.754144966112236,
|
183 |
-
('sub_no-matrix', 1): 18.909088991066888,
|
184 |
-
('sub_no-matrix', 2): 57.753410746636746,
|
185 |
-
('sub_no-matrix', 3): 28.677667873674363,
|
186 |
-
('sub_no-matrix', 4): 51.9941480032352,
|
187 |
-
('sub_no-matrix', 5): 5.033266273930268},
|
188 |
-
{('no-sub_matrix', 1): 14.859413855165633,
|
189 |
-
('no-sub_matrix', 2): 34.54519231993284,
|
190 |
-
('no-sub_matrix', 3): 24.26528519671309,
|
191 |
-
('no-sub_matrix', 4): 35.42343514121054,
|
192 |
-
('no-sub_matrix', 5): 55.85308623165151,
|
193 |
-
('no-sub_no-matrix', 1): 14.859413855165633,
|
194 |
-
('no-sub_no-matrix', 2): 34.54519231993284,
|
195 |
-
('no-sub_no-matrix', 3): 24.26528519671309,
|
196 |
-
('no-sub_no-matrix', 4): 35.42343514121054,
|
197 |
-
('no-sub_no-matrix', 5): 2.3309861205259734,
|
198 |
-
('sub_matrix', 1): 17.053809634549854,
|
199 |
-
('sub_matrix', 2): 33.66637542056656,
|
200 |
-
('sub_matrix', 3): 23.26181234829638,
|
201 |
-
('sub_matrix', 4): 35.61438567264568,
|
202 |
-
('sub_matrix', 5): 48.48551986050014,
|
203 |
-
('sub_no-matrix', 1): 17.053809634549854,
|
204 |
-
('sub_no-matrix', 2): 33.66637542056656,
|
205 |
-
('sub_no-matrix', 3): 23.26181234829638,
|
206 |
-
('sub_no-matrix', 4): 35.61438704850689,
|
207 |
-
('sub_no-matrix', 5): 2.969309360231736},
|
208 |
-
{('no-sub_matrix', 1): 13.708973748402064,
|
209 |
-
('no-sub_matrix', 2): 31.147590264691182,
|
210 |
-
('no-sub_matrix', 3): 30.495597241955565,
|
211 |
-
('no-sub_matrix', 4): 34.65164493728535,
|
212 |
-
('no-sub_matrix', 5): 35.87510990950117,
|
213 |
-
('no-sub_no-matrix', 1): 13.708973748402064,
|
214 |
-
('no-sub_no-matrix', 2): 31.147590264691182,
|
215 |
-
('no-sub_no-matrix', 3): 30.495597241955565,
|
216 |
-
('no-sub_no-matrix', 4): 34.65164493728535,
|
217 |
-
('no-sub_no-matrix', 5): 3.232032121481573,
|
218 |
-
('sub_matrix', 1): 17.681722076468287,
|
219 |
-
('sub_matrix', 2): 33.77225997922327,
|
220 |
-
('sub_matrix', 3): 29.435808932487806,
|
221 |
-
('sub_matrix', 4): 34.354368969668016,
|
222 |
-
('sub_matrix', 5): 20.802733205442486,
|
223 |
-
('sub_no-matrix', 1): 17.681722076468287,
|
224 |
-
('sub_no-matrix', 2): 33.77225997922327,
|
225 |
-
('sub_no-matrix', 3): 29.435808932487806,
|
226 |
-
('sub_no-matrix', 4): 34.354368969668016,
|
227 |
-
('sub_no-matrix', 5): 3.7902066303710424},
|
228 |
-
{('no-sub_matrix', 1): 15.72185319065555,
|
229 |
-
('no-sub_matrix', 2): 45.25539814380218,
|
230 |
-
('no-sub_matrix', 3): 24.94273362957689,
|
231 |
-
('no-sub_matrix', 4): 40.81704901026569,
|
232 |
-
('no-sub_matrix', 5): 42.898794519499596,
|
233 |
-
('no-sub_no-matrix', 1): 15.72185319065555,
|
234 |
-
('no-sub_no-matrix', 2): 45.25539814380218,
|
235 |
-
('no-sub_no-matrix', 3): 24.94273362957689,
|
236 |
-
('no-sub_no-matrix', 4): 40.81704901026569,
|
237 |
-
('no-sub_no-matrix', 5): 2.6826901255924644,
|
238 |
-
('sub_matrix', 1): 17.565795106862403,
|
239 |
-
('sub_matrix', 2): 46.9371803702329,
|
240 |
-
('sub_matrix', 3): 23.887805807796486,
|
241 |
-
('sub_matrix', 4): 39.058599411828766,
|
242 |
-
('sub_matrix', 5): 32.234453544910295,
|
243 |
-
('sub_no-matrix', 1): 17.565795106862403,
|
244 |
-
('sub_no-matrix', 2): 46.9371803702329,
|
245 |
-
('sub_no-matrix', 3): 23.887805807796486,
|
246 |
-
('sub_no-matrix', 4): 39.058599411828766,
|
247 |
-
('sub_no-matrix', 5): 4.214674259243127},
|
248 |
-
{('no-sub_matrix', 1): 13.910878628792588,
|
249 |
-
('no-sub_matrix', 2): 33.45626834359109,
|
250 |
-
('no-sub_matrix', 3): 16.127584513594687,
|
251 |
-
('no-sub_matrix', 4): 32.59623120264939,
|
252 |
-
('no-sub_matrix', 5): 29.87568851789407,
|
253 |
-
('no-sub_no-matrix', 1): 13.910878628792588,
|
254 |
-
('no-sub_no-matrix', 2): 33.45626834359109,
|
255 |
-
('no-sub_no-matrix', 3): 16.127584513594687,
|
256 |
-
('no-sub_no-matrix', 4): 32.59623120264939,
|
257 |
-
('no-sub_no-matrix', 5): 2.3891779982892625,
|
258 |
-
('sub_matrix', 1): 17.18981661053988,
|
259 |
-
('sub_matrix', 2): 36.38883326650068,
|
260 |
-
('sub_matrix', 3): 13.081088737716442,
|
261 |
-
('sub_matrix', 4): 33.419732612590224,
|
262 |
-
('sub_matrix', 5): 22.665485632721676,
|
263 |
-
('sub_no-matrix', 1): 17.18981661053988,
|
264 |
-
('sub_no-matrix', 2): 36.38883326650068,
|
265 |
-
('sub_no-matrix', 3): 13.081088737716442,
|
266 |
-
('sub_no-matrix', 4): 33.419732612590224,
|
267 |
-
('sub_no-matrix', 5): 6.155199912348024},
|
268 |
-
{('no-sub_matrix', 1): 18.196771699177763,
|
269 |
-
('no-sub_matrix', 2): 35.624058750852136,
|
270 |
-
('no-sub_matrix', 3): 23.746554392851053,
|
271 |
-
('no-sub_matrix', 4): 29.44669921790574,
|
272 |
-
('no-sub_matrix', 5): 39.72412918901379,
|
273 |
-
('no-sub_no-matrix', 1): 18.196771699177763,
|
274 |
-
('no-sub_no-matrix', 2): 35.624058750852136,
|
275 |
-
('no-sub_no-matrix', 3): 23.746554392851053,
|
276 |
-
('no-sub_no-matrix', 4): 29.44669921790574,
|
277 |
-
('no-sub_no-matrix', 5): 2.870123353843486,
|
278 |
-
('sub_matrix', 1): 20.38619930823735,
|
279 |
-
('sub_matrix', 2): 36.29781144853154,
|
280 |
-
('sub_matrix', 3): 22.13637404741934,
|
281 |
-
('sub_matrix', 4): 29.68729899086184,
|
282 |
-
('sub_matrix', 5): 36.993790238103884,
|
283 |
-
('sub_no-matrix', 1): 20.38619930823735,
|
284 |
-
('sub_no-matrix', 2): 36.29781144853154,
|
285 |
-
('sub_no-matrix', 3): 22.13637404741934,
|
286 |
-
('sub_no-matrix', 4): 29.68729899086184,
|
287 |
-
('sub_no-matrix', 5): 7.650303570399713},
|
288 |
-
{('no-sub_matrix', 1): 11.992867568477442,
|
289 |
-
('no-sub_matrix', 2): 26.44083030170154,
|
290 |
-
('no-sub_matrix', 3): 27.574921221726136,
|
291 |
-
('no-sub_matrix', 4): 28.94213565689118,
|
292 |
-
('no-sub_matrix', 5): 46.973469397495556,
|
293 |
-
('no-sub_no-matrix', 1): 11.992867568477442,
|
294 |
-
('no-sub_no-matrix', 2): 26.44083030170154,
|
295 |
-
('no-sub_no-matrix', 3): 27.574921221726136,
|
296 |
-
('no-sub_no-matrix', 4): 28.94213565689118,
|
297 |
-
('no-sub_no-matrix', 5): 3.354326576753004,
|
298 |
-
('sub_matrix', 1): 14.434047100994839,
|
299 |
-
('sub_matrix', 2): 26.76571524620116,
|
300 |
-
('sub_matrix', 3): 25.83488399989926,
|
301 |
-
('sub_matrix', 4): 30.263621195061678,
|
302 |
-
('sub_matrix', 5): 36.822532494114455,
|
303 |
-
('sub_no-matrix', 1): 14.434047100994839,
|
304 |
-
('sub_no-matrix', 2): 26.76571524620116,
|
305 |
-
('sub_no-matrix', 3): 25.83488399989926,
|
306 |
-
('sub_no-matrix', 4): 30.263621195061678,
|
307 |
-
('sub_no-matrix', 5): 6.748976893757906},
|
308 |
-
{('no-sub_matrix', 1): 16.27614914680276,
|
309 |
-
('no-sub_matrix', 2): 41.35282905624703,
|
310 |
-
('no-sub_matrix', 3): 25.173115913245226,
|
311 |
-
('no-sub_matrix', 4): 52.876981987369014,
|
312 |
-
('no-sub_matrix', 5): 49.49767321075167,
|
313 |
-
('no-sub_no-matrix', 1): 16.27614914680276,
|
314 |
-
('no-sub_no-matrix', 2): 41.35282905624703,
|
315 |
-
('no-sub_no-matrix', 3): 25.173115913245226,
|
316 |
-
('no-sub_no-matrix', 4): 52.876981987369014,
|
317 |
-
('no-sub_no-matrix', 5): 1.5962803636236758,
|
318 |
-
('sub_matrix', 1): 18.735912436641787,
|
319 |
-
('sub_matrix', 2): 43.36213985849511,
|
320 |
-
('sub_matrix', 3): 24.582800598631913,
|
321 |
-
('sub_matrix', 4): 53.1616607417586,
|
322 |
-
('sub_matrix', 5): 41.2664433745972,
|
323 |
-
('sub_no-matrix', 1): 18.735912436641787,
|
324 |
-
('sub_no-matrix', 2): 43.36213985849511,
|
325 |
-
('sub_no-matrix', 3): 24.582800598631913,
|
326 |
-
('sub_no-matrix', 4): 53.16165799003619,
|
327 |
-
('sub_no-matrix', 5): 6.4917878462822305},
|
328 |
-
{('no-sub_matrix', 1): 14.036280122634507,
|
329 |
-
('no-sub_matrix', 2): 53.72802368862095,
|
330 |
-
('no-sub_matrix', 3): 18.940766131564004,
|
331 |
-
('no-sub_matrix', 4): 40.74964840745327,
|
332 |
-
('no-sub_matrix', 5): 39.57008490907742,
|
333 |
-
('no-sub_no-matrix', 1): 14.036280122634507,
|
334 |
-
('no-sub_no-matrix', 2): 53.72802368862095,
|
335 |
-
('no-sub_no-matrix', 3): 18.940766131564004,
|
336 |
-
('no-sub_no-matrix', 4): 40.74964840745327,
|
337 |
-
('no-sub_no-matrix', 5): 2.1275557540222967,
|
338 |
-
('sub_matrix', 1): 19.641722357026286,
|
339 |
-
('sub_matrix', 2): 52.709120728751486,
|
340 |
-
('sub_matrix', 3): 17.976257844509426,
|
341 |
-
('sub_matrix', 4): 42.51851542500959,
|
342 |
-
('sub_matrix', 5): 28.25018664655579,
|
343 |
-
('sub_no-matrix', 1): 19.641722357026286,
|
344 |
-
('sub_no-matrix', 2): 52.709120728751486,
|
345 |
-
('sub_no-matrix', 3): 17.976257844509426,
|
346 |
-
('sub_no-matrix', 4): 42.51851267328718,
|
347 |
-
('sub_no-matrix', 5): 5.409622788119386},
|
348 |
-
{('no-sub_matrix', 1): 16.961927903326398,
|
349 |
-
('no-sub_matrix', 2): 38.5455951142925,
|
350 |
-
('no-sub_matrix', 3): 25.122316709729276,
|
351 |
-
('no-sub_matrix', 4): 35.90131439006518,
|
352 |
-
('no-sub_matrix', 5): 41.65886977570029,
|
353 |
-
('no-sub_no-matrix', 1): 16.961927903326398,
|
354 |
-
('no-sub_no-matrix', 2): 38.5455951142925,
|
355 |
-
('no-sub_no-matrix', 3): 25.122316709729276,
|
356 |
-
('no-sub_no-matrix', 4): 35.90131439006518,
|
357 |
-
('no-sub_no-matrix', 5): 3.2679255886472447,
|
358 |
-
('sub_matrix', 1): 20.247934372024154,
|
359 |
-
('sub_matrix', 2): 40.408716019775625,
|
360 |
-
('sub_matrix', 3): 23.782735071043668,
|
361 |
-
('sub_matrix', 4): 37.00513584758997,
|
362 |
-
('sub_matrix', 5): 29.22700479607527,
|
363 |
-
('sub_no-matrix', 1): 20.247934372024154,
|
364 |
-
('sub_no-matrix', 2): 40.408716019775625,
|
365 |
-
('sub_no-matrix', 3): 23.782735071043668,
|
366 |
-
('sub_no-matrix', 4): 37.00513584758997,
|
367 |
-
('sub_no-matrix', 5): 4.780011845541033},
|
368 |
-
{('no-sub_matrix', 1): 12.109815771064152,
|
369 |
-
('no-sub_matrix', 2): 38.32406752938649,
|
370 |
-
('no-sub_matrix', 3): 25.987801084044044,
|
371 |
-
('no-sub_matrix', 4): 40.40950903177875,
|
372 |
-
('no-sub_matrix', 5): 52.86522525335603,
|
373 |
-
('no-sub_no-matrix', 1): 12.109815771064152,
|
374 |
-
('no-sub_no-matrix', 2): 38.32406752938649,
|
375 |
-
('no-sub_no-matrix', 3): 25.987801084044044,
|
376 |
-
('no-sub_no-matrix', 4): 40.40950903177875,
|
377 |
-
('no-sub_no-matrix', 5): 3.61917194787979,
|
378 |
-
('sub_matrix', 1): 15.130341564722832,
|
379 |
-
('sub_matrix', 2): 37.89719334728088,
|
380 |
-
('sub_matrix', 3): 24.65681032273433,
|
381 |
-
('sub_matrix', 4): 40.731610867030774,
|
382 |
-
('sub_matrix', 5): 37.566910985257906,
|
383 |
-
('sub_no-matrix', 1): 15.130341564722832,
|
384 |
-
('sub_no-matrix', 2): 37.89719334728088,
|
385 |
-
('sub_no-matrix', 3): 24.65681032273433,
|
386 |
-
('sub_no-matrix', 4): 40.731610867030774,
|
387 |
-
('sub_no-matrix', 5): 9.39736249989602},
|
388 |
-
{('no-sub_matrix', 1): 16.25058564557851,
|
389 |
-
('no-sub_matrix', 2): 37.20405682898803,
|
390 |
-
('no-sub_matrix', 3): 30.5107090995129,
|
391 |
-
('no-sub_matrix', 4): 44.537084655292894,
|
392 |
-
('no-sub_matrix', 5): 46.50046620075818,
|
393 |
-
('no-sub_no-matrix', 1): 16.25058564557851,
|
394 |
-
('no-sub_no-matrix', 2): 37.20405682898803,
|
395 |
-
('no-sub_no-matrix', 3): 30.5107090995129,
|
396 |
-
('no-sub_no-matrix', 4): 44.537084655292894,
|
397 |
-
('no-sub_no-matrix', 5): 1.8752506698658238,
|
398 |
-
('sub_matrix', 1): 18.440281483079957,
|
399 |
-
('sub_matrix', 2): 38.54769605435544,
|
400 |
-
('sub_matrix', 3): 30.510800250317864,
|
401 |
-
('sub_matrix', 4): 44.99740645329493,
|
402 |
-
('sub_matrix', 5): 39.55738177603457,
|
403 |
-
('sub_no-matrix', 1): 18.440281483079957,
|
404 |
-
('sub_no-matrix', 2): 38.54769605435544,
|
405 |
-
('sub_no-matrix', 3): 30.510800250317864,
|
406 |
-
('sub_no-matrix', 4): 44.99740645329493,
|
407 |
-
('sub_no-matrix', 5): 2.6233048602148386},
|
408 |
-
{('no-sub_matrix', 1): 16.324447378609865,
|
409 |
-
('no-sub_matrix', 2): 30.87308462806543,
|
410 |
-
('no-sub_matrix', 3): 22.765564836381643,
|
411 |
-
('no-sub_matrix', 4): 38.337445027901204,
|
412 |
-
('no-sub_matrix', 5): 40.98815076599078,
|
413 |
-
('no-sub_no-matrix', 1): 16.324447378609865,
|
414 |
-
('no-sub_no-matrix', 2): 30.87308462806543,
|
415 |
-
('no-sub_no-matrix', 3): 22.765564836381643,
|
416 |
-
('no-sub_no-matrix', 4): 38.337445027901204,
|
417 |
-
('no-sub_no-matrix', 5): 1.4796406979126138,
|
418 |
-
('sub_matrix', 1): 17.9623592385626,
|
419 |
-
('sub_matrix', 2): 32.36568198294609,
|
420 |
-
('sub_matrix', 3): 22.438215466486483,
|
421 |
-
('sub_matrix', 4): 40.900713840387546,
|
422 |
-
('sub_matrix', 5): 33.396627340011634,
|
423 |
-
('sub_no-matrix', 1): 17.9623592385626,
|
424 |
-
('sub_no-matrix', 2): 32.36568198294609,
|
425 |
-
('sub_no-matrix', 3): 22.438215466486483,
|
426 |
-
('sub_no-matrix', 4): 40.900713840387546,
|
427 |
-
('sub_no-matrix', 5): 6.609518913895668},
|
428 |
-
{('no-sub_matrix', 1): 14.033258731424148,
|
429 |
-
('no-sub_matrix', 2): 28.37206528002418,
|
430 |
-
('no-sub_matrix', 3): 27.043658386061033,
|
431 |
-
('no-sub_matrix', 4): 36.167049513436204,
|
432 |
-
('no-sub_matrix', 5): 52.280797076864395,
|
433 |
-
('no-sub_no-matrix', 1): 14.033258731424148,
|
434 |
-
('no-sub_no-matrix', 2): 28.37206528002418,
|
435 |
-
('no-sub_no-matrix', 3): 27.043658386061033,
|
436 |
-
('no-sub_no-matrix', 4): 36.167049513436204,
|
437 |
-
('no-sub_no-matrix', 5): 1.9358795417918389,
|
438 |
-
('sub_matrix', 1): 16.606623097498794,
|
439 |
-
('sub_matrix', 2): 29.98729916366884,
|
440 |
-
('sub_matrix', 3): 24.737985875967603,
|
441 |
-
('sub_matrix', 4): 34.93154214402433,
|
442 |
-
('sub_matrix', 5): 42.35241303296243,
|
443 |
-
('sub_no-matrix', 1): 16.606623097498794,
|
444 |
-
('sub_no-matrix', 2): 29.98729916366884,
|
445 |
-
('sub_no-matrix', 3): 24.737985875967603,
|
446 |
-
('sub_no-matrix', 4): 34.931551775052775,
|
447 |
-
('sub_no-matrix', 5): 7.151971456773863},
|
448 |
-
{('no-sub_matrix', 1): 10.482293039084738,
|
449 |
-
('no-sub_matrix', 2): 52.67861788579445,
|
450 |
-
('no-sub_matrix', 3): 21.665543335527666,
|
451 |
-
('no-sub_matrix', 4): 23.53727708917033,
|
452 |
-
('no-sub_matrix', 5): 32.2645584918966,
|
453 |
-
('no-sub_no-matrix', 1): 10.482293039084738,
|
454 |
-
('no-sub_no-matrix', 2): 52.67861788579445,
|
455 |
-
('no-sub_no-matrix', 3): 21.665543335527666,
|
456 |
-
('no-sub_no-matrix', 4): 23.53727708917033,
|
457 |
-
('no-sub_no-matrix', 5): 2.5207572809328243,
|
458 |
-
('sub_matrix', 1): 11.523882918360123,
|
459 |
-
('sub_matrix', 2): 57.336257883871156,
|
460 |
-
('sub_matrix', 3): 21.647716645835132,
|
461 |
-
('sub_matrix', 4): 23.491483569694733,
|
462 |
-
('sub_matrix', 5): 24.264706351480406,
|
463 |
-
('sub_no-matrix', 1): 11.523882918360123,
|
464 |
-
('sub_no-matrix', 2): 57.336257883871156,
|
465 |
-
('sub_no-matrix', 3): 21.647716645835132,
|
466 |
-
('sub_no-matrix', 4): 23.491462243846026,
|
467 |
-
('sub_no-matrix', 5): 9.714244661694366},
|
468 |
-
{('no-sub_matrix', 1): 11.992867568477442,
|
469 |
-
('no-sub_matrix', 2): 28.861638231250264,
|
470 |
-
('no-sub_matrix', 3): 24.222607873884137,
|
471 |
-
('no-sub_matrix', 4): 41.28280460012173,
|
472 |
-
('no-sub_matrix', 5): 56.6084264455065,
|
473 |
-
('no-sub_no-matrix', 1): 11.992867568477442,
|
474 |
-
('no-sub_no-matrix', 2): 28.861638231250264,
|
475 |
-
('no-sub_no-matrix', 3): 24.222607873884137,
|
476 |
-
('no-sub_no-matrix', 4): 41.28280460012173,
|
477 |
-
('no-sub_no-matrix', 5): 2.4980576348107437,
|
478 |
-
('sub_matrix', 1): 14.531057698832324,
|
479 |
-
('sub_matrix', 2): 31.280393934821902,
|
480 |
-
('sub_matrix', 3): 20.756528260470358,
|
481 |
-
('sub_matrix', 4): 42.15937712589425,
|
482 |
-
('sub_matrix', 5): 52.45767194621365,
|
483 |
-
('sub_no-matrix', 1): 14.531057698832324,
|
484 |
-
('sub_no-matrix', 2): 31.280393934821902,
|
485 |
-
('sub_no-matrix', 3): 20.756528260470358,
|
486 |
-
('sub_no-matrix', 4): 42.15937712589425,
|
487 |
-
('sub_no-matrix', 5): 4.819862633503057}]
|
488 |
-
|
489 |
-
|
490 |
-
def test_gpt_subordination_region_totals():
|
491 |
-
"""
|
492 |
-
Check region-level surprisals against the original syntaxgym-core
|
493 |
-
implementation, using the same underlying `gpt2` model.
|
494 |
-
"""
|
495 |
-
reference = ... # TODO
|
496 |
-
|
497 |
-
# TODO work out references
|
498 |
-
dataset = datasets.load_dataset("./syntaxgym.py", "subordination_src-src")
|
499 |
-
metric = datasets.load_metric("./syntaxgym.py")
|
500 |
-
result = metric.compute(suite=dataset["test"], model_id="gpt2")
|
501 |
-
|
502 |
-
from pprint import pprint
|
503 |
-
pprint(result["region_totals"][0])
|
504 |
-
pprint(GPT2_SUBORDINATION_SRC_REFERENCE[0])
|
505 |
-
|
506 |
-
keys = result["region_totals"][0].keys()
|
507 |
-
assert set(keys) == set(GPT2_SUBORDINATION_SRC_REFERENCE[0].keys())
|
508 |
-
|
509 |
-
result_ndarray = np.concatenate([np.array([region_totals[key] for key in keys])
|
510 |
-
for region_totals in result["region_totals"]])
|
511 |
-
reference_ndarray = np.concatenate([np.array([region_totals[key] for key in keys])
|
512 |
-
for region_totals in GPT2_SUBORDINATION_SRC_REFERENCE])
|
513 |
-
pprint(sorted(zip(keys, np.abs(result_ndarray - reference_ndarray)),
|
514 |
-
key=lambda x: -x[1]))
|
515 |
-
np.testing.assert_allclose(result_ndarray, reference_ndarray, atol=1e-3)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|