Spaces:
Running
Running
Update app.py
Browse files
app.py
CHANGED
@@ -1,23 +1,21 @@
|
|
1 |
#!/usr/bin/env python3
|
2 |
"""
|
3 |
-
what_comes_next.py – Hugging
|
4 |
-
A
|
5 |
-
|
6 |
-
|
7 |
-
|
8 |
-
|
9 |
-
|
10 |
-
|
11 |
-
|
12 |
-
|
13 |
-
|
14 |
-
•
|
15 |
-
`data.json` (JSON‑Lines). When the Oracle finally finishes, offline evaluation can
|
16 |
-
score the guesses against the final text.
|
17 |
-
|
18 |
-
The game then moves on to the next prompt and the cycle repeats.
|
19 |
"""
|
20 |
|
|
|
|
|
21 |
import os
|
22 |
import json
|
23 |
import time
|
@@ -25,47 +23,45 @@ import random
|
|
25 |
import threading
|
26 |
import logging
|
27 |
from datetime import datetime, timezone
|
28 |
-
from pathlib import Path
|
29 |
from typing import Dict, Any
|
30 |
|
31 |
import torch
|
32 |
-
from transformers import AutoTokenizer, AutoModelForCausalLM, TextIteratorStreamer
|
33 |
import gradio as gr
|
|
|
34 |
|
35 |
###############################################################################
|
36 |
-
#
|
37 |
###############################################################################
|
38 |
-
MODEL_NAME = "meta-llama/Llama-3.1-8B-Instruct"
|
39 |
-
PROMPTS_PATH = "
|
40 |
-
STATE_PATH = "current_state.json"
|
41 |
-
DATA_PATH = "data.json"
|
42 |
-
|
43 |
-
|
44 |
-
|
45 |
-
|
46 |
-
|
|
|
47 |
###############################################################################
|
48 |
|
49 |
logging.basicConfig(format="[%(asctime)s] %(levelname)s: %(message)s", level=logging.INFO)
|
50 |
-
log = logging.getLogger("what
|
51 |
-
|
52 |
-
lock = threading.Lock() # global file/variable lock
|
53 |
|
54 |
-
|
55 |
-
#
|
56 |
-
|
57 |
|
58 |
def _read_json(path: str, default: Any):
|
59 |
try:
|
60 |
-
with open(path, "r", encoding="utf
|
61 |
return json.load(f)
|
62 |
except FileNotFoundError:
|
63 |
return default
|
64 |
|
65 |
|
66 |
-
def
|
67 |
tmp = f"{path}.tmp"
|
68 |
-
with open(tmp, "w", encoding="utf
|
69 |
json.dump(obj, f, ensure_ascii=False, indent=2)
|
70 |
os.replace(tmp, path)
|
71 |
|
@@ -73,170 +69,156 @@ def _write_json(path: str, obj: Any):
|
|
73 |
def load_prompts() -> list[str]:
|
74 |
if not os.path.exists(PROMPTS_PATH):
|
75 |
raise FileNotFoundError(f"Missing {PROMPTS_PATH}. Please add 100 prompts.")
|
76 |
-
with open(PROMPTS_PATH, "r", encoding="utf
|
77 |
-
|
|
|
|
|
|
|
78 |
|
79 |
-
|
80 |
-
|
81 |
-
|
82 |
-
|
83 |
-
# Model loading (FP32 ‑ CPU) #
|
84 |
-
# --------------------------------------------------------------------------- #
|
85 |
-
log.info("Loading Llama‑3.1‑8B‑Instruct in FP32 on CPU (this is *slow*) …")
|
86 |
|
87 |
tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
|
88 |
model = AutoModelForCausalLM.from_pretrained(
|
89 |
MODEL_NAME,
|
90 |
torch_dtype=torch.float32,
|
91 |
-
device_map={"": "cpu"}, # force CPU
|
92 |
)
|
93 |
model.eval()
|
94 |
-
log.info("Model
|
95 |
|
96 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
97 |
# Oracle generation thread #
|
98 |
-
|
99 |
|
100 |
-
def
|
101 |
-
"""Return existing state or create a
|
102 |
state = _read_json(STATE_PATH, {})
|
103 |
-
if state.get("finished"
|
104 |
-
state = {} # finished, start new prompt
|
105 |
-
if not state:
|
106 |
prompt_idx = random.randrange(len(prompts))
|
107 |
-
prompt = prompts[prompt_idx]
|
108 |
state = {
|
109 |
"prompt_idx": prompt_idx,
|
110 |
-
"prompt":
|
111 |
-
"generated": "",
|
|
|
112 |
"start_time": time.time(),
|
113 |
-
"finished": False
|
114 |
-
"tokens_done": 0
|
115 |
}
|
116 |
-
|
117 |
-
log.info(f"
|
118 |
return state
|
119 |
|
120 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
121 |
def oracle_loop():
|
122 |
-
"""Continuously extend the Oracle’s text by one token every SECS_BETWEEN_TOKENS."""
|
123 |
while True:
|
124 |
with lock:
|
125 |
-
state =
|
126 |
-
|
127 |
-
|
128 |
-
|
129 |
-
|
130 |
-
|
131 |
-
|
132 |
-
|
133 |
-
|
134 |
-
#
|
135 |
-
full_input = prompt_text + generated_text
|
136 |
-
input_ids = tokenizer(full_input, return_tensors="pt", truncation=True, max_length=MAX_CONTEXT_TOKENS).input_ids
|
137 |
-
|
138 |
-
# Generate ONE token
|
139 |
with torch.no_grad():
|
140 |
-
|
141 |
input_ids,
|
142 |
max_new_tokens=1,
|
143 |
do_sample=True,
|
144 |
temperature=TEMPERATURE,
|
145 |
top_p=TOP_P,
|
146 |
)
|
147 |
-
|
148 |
-
next_token_text = tokenizer.decode(next_token_id, skip_special_tokens=True, clean_up_tokenization_spaces=False)
|
149 |
|
150 |
with lock:
|
151 |
-
|
152 |
-
state["generated"] += next_token_text
|
153 |
state["tokens_done"] += 1
|
154 |
if state["tokens_done"] >= TOKENS_PER_PROMPT:
|
155 |
state["finished"] = True
|
156 |
-
log.info("Prompt
|
157 |
-
|
158 |
-
time.sleep(SECS_BETWEEN_TOKENS)
|
159 |
-
|
160 |
|
161 |
threading.Thread(target=oracle_loop, daemon=True).start()
|
162 |
|
163 |
-
|
164 |
-
# Gradio
|
165 |
-
|
166 |
|
167 |
-
def
|
168 |
-
|
169 |
-
|
170 |
-
|
171 |
-
return
|
172 |
|
173 |
|
174 |
-
def
|
175 |
-
|
176 |
-
|
177 |
-
if not
|
178 |
-
return
|
179 |
-
return {
|
180 |
-
"prompt": state["prompt"],
|
181 |
-
"generated": state["generated"],
|
182 |
-
"elapsed": human_readable_elapsed(state["start_time"])
|
183 |
-
}
|
184 |
|
|
|
|
|
|
|
185 |
|
186 |
-
def record_guess(full_guess: str, idea_guess: str):
|
187 |
-
state = get_current_state()
|
188 |
-
guess_text = full_guess.strip() or idea_guess.strip()
|
189 |
-
if not guess_text:
|
190 |
-
return gr.update(value="⚠️ Please enter a guess in one of the boxes …"), gr.update()
|
191 |
-
guess_type = "full" if full_guess.strip() else "idea"
|
192 |
record = {
|
193 |
"timestamp": datetime.now(timezone.utc).isoformat(),
|
194 |
-
"prompt":
|
195 |
-
"point
|
196 |
-
"response
|
197 |
-
"user
|
198 |
-
"guess
|
199 |
}
|
200 |
-
# Append to JSONL (data.json)
|
201 |
with lock:
|
202 |
-
with open(DATA_PATH, "a", encoding="utf
|
203 |
f.write(json.dumps(record, ensure_ascii=False) + "\n")
|
204 |
-
log.info(f"
|
205 |
-
return gr.update(value="✅
|
206 |
|
207 |
|
208 |
-
with gr.Blocks(title="What
|
209 |
-
gr.Markdown("""#
|
210 |
-
|
211 |
-
|
212 |
-
|
213 |
|
214 |
-
|
215 |
-
|
216 |
-
|
217 |
-
elapsed_box = gr.Textbox(label="⏱️ Elapsed", interactive=False)
|
218 |
-
|
219 |
-
### Guess inputs
|
220 |
-
gr.Markdown("**Make your prediction:** Fill **either** the exact continuation *or* a general idea.")
|
221 |
-
with gr.Row():
|
222 |
-
full_guess = gr.Textbox(label="🧠 Exact continuation (full)")
|
223 |
-
idea_guess = gr.Textbox(label="💡 General idea")
|
224 |
-
submit_btn = gr.Button("Submit Guess")
|
225 |
-
status_msg = gr.Textbox(label="Status", interactive=False)
|
226 |
|
227 |
-
|
228 |
-
refresh_btn = gr.Button("🔄 Refresh Oracle progress")
|
229 |
|
230 |
-
|
231 |
-
|
232 |
-
|
233 |
-
|
234 |
-
|
235 |
-
|
236 |
-
|
237 |
-
|
238 |
-
|
239 |
-
|
|
|
|
|
|
|
240 |
|
241 |
if __name__ == "__main__":
|
242 |
demo.launch(server_name="0.0.0.0", server_port=7860, show_error=True)
|
|
|
1 |
#!/usr/bin/env python3
|
2 |
"""
|
3 |
+
what_comes_next.py – Hugging Face Space implementation of **What Comes Next**
|
4 |
+
A global, slow-burn guessing game powered by Llama-3.1-8B-Instruct (FP32, CPU-only).
|
5 |
+
|
6 |
+
HOW IT WORKS
|
7 |
+
============
|
8 |
+
• One shared model generates a single, very long completion (≈2 k tokens) for a chosen
|
9 |
+
prompt in *full precision* on CPU. One token is sampled every ~15 s, so a prompt
|
10 |
+
unfolds for roughly 10 hours. All visitors see the same progress in real-time.
|
11 |
+
• Players read the partial output and may submit **either**
|
12 |
+
🧠 Exact continuation (full guess) **or** 💡 General idea (summary guess).
|
13 |
+
• Each guess is appended to `data.json` with prompt, Oracle progress, timestamp & type.
|
14 |
+
• Offline scoring (not included here) can later measure similarity vs the final text.
|
|
|
|
|
|
|
|
|
15 |
"""
|
16 |
|
17 |
+
from __future__ import annotations
|
18 |
+
|
19 |
import os
|
20 |
import json
|
21 |
import time
|
|
|
23 |
import threading
|
24 |
import logging
|
25 |
from datetime import datetime, timezone
|
|
|
26 |
from typing import Dict, Any
|
27 |
|
28 |
import torch
|
|
|
29 |
import gradio as gr
|
30 |
+
from transformers import AutoTokenizer, AutoModelForCausalLM
|
31 |
|
32 |
###############################################################################
|
33 |
+
# Configuration #
|
34 |
###############################################################################
|
35 |
+
MODEL_NAME = "meta-llama/Llama-3.1-8B-Instruct" # full-precision model
|
36 |
+
PROMPTS_PATH = "full_prompts.json" # 100 full prompts
|
37 |
+
STATE_PATH = "current_state.json" # persistent Oracle state
|
38 |
+
DATA_PATH = "data.json" # JSONL log of guesses
|
39 |
+
|
40 |
+
TOKENS_PER_PROMPT = 2048 # stop after N generated tokens
|
41 |
+
SECS_BETWEEN_TOKENS = 15 # ~10 h per prompt
|
42 |
+
TEMPERATURE = 0.9 # higher creativity, as requested
|
43 |
+
TOP_P = 0.95 # nucleus sampling
|
44 |
+
MAX_CONTEXT_TOKENS = 8192 # safety cap
|
45 |
###############################################################################
|
46 |
|
47 |
logging.basicConfig(format="[%(asctime)s] %(levelname)s: %(message)s", level=logging.INFO)
|
48 |
+
log = logging.getLogger("what-comes-next")
|
|
|
|
|
49 |
|
50 |
+
###############################################################################
|
51 |
+
# Utility helpers #
|
52 |
+
###############################################################################
|
53 |
|
54 |
def _read_json(path: str, default: Any):
|
55 |
try:
|
56 |
+
with open(path, "r", encoding="utf-8") as f:
|
57 |
return json.load(f)
|
58 |
except FileNotFoundError:
|
59 |
return default
|
60 |
|
61 |
|
62 |
+
def _atomic_write(path: str, obj: Any):
|
63 |
tmp = f"{path}.tmp"
|
64 |
+
with open(tmp, "w", encoding="utf-8") as f:
|
65 |
json.dump(obj, f, ensure_ascii=False, indent=2)
|
66 |
os.replace(tmp, path)
|
67 |
|
|
|
69 |
def load_prompts() -> list[str]:
|
70 |
if not os.path.exists(PROMPTS_PATH):
|
71 |
raise FileNotFoundError(f"Missing {PROMPTS_PATH}. Please add 100 prompts.")
|
72 |
+
with open(PROMPTS_PATH, "r", encoding="utf-8") as f:
|
73 |
+
prompts = json.load(f)
|
74 |
+
if not isinstance(prompts, list) or not prompts:
|
75 |
+
raise ValueError("full_prompts.json must be a non-empty JSON array of strings")
|
76 |
+
return prompts
|
77 |
|
78 |
+
###############################################################################
|
79 |
+
# Model loading #
|
80 |
+
###############################################################################
|
81 |
+
log.info("Loading Llama-3.1-8B-Instruct (FP32 CPU-only)… this can take a while.")
|
|
|
|
|
|
|
82 |
|
83 |
tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
|
84 |
model = AutoModelForCausalLM.from_pretrained(
|
85 |
MODEL_NAME,
|
86 |
torch_dtype=torch.float32,
|
87 |
+
device_map={"": "cpu"}, # force CPU placement
|
88 |
)
|
89 |
model.eval()
|
90 |
+
log.info("Model ready – Oracle awakened.")
|
91 |
|
92 |
+
###############################################################################
|
93 |
+
# Global state #
|
94 |
+
###############################################################################
|
95 |
+
lock = threading.Lock() # guard state + files
|
96 |
+
prompts = load_prompts() # list of 100 strings
|
97 |
+
|
98 |
+
###############################################################################
|
99 |
# Oracle generation thread #
|
100 |
+
###############################################################################
|
101 |
|
102 |
+
def _init_state() -> Dict[str, Any]:
|
103 |
+
"""Return existing state or create a fresh one if none/finished."""
|
104 |
state = _read_json(STATE_PATH, {})
|
105 |
+
if not state or state.get("finished"):
|
|
|
|
|
106 |
prompt_idx = random.randrange(len(prompts))
|
|
|
107 |
state = {
|
108 |
"prompt_idx": prompt_idx,
|
109 |
+
"prompt": prompts[prompt_idx],
|
110 |
+
"generated": "", # text so far
|
111 |
+
"tokens_done": 0,
|
112 |
"start_time": time.time(),
|
113 |
+
"finished": False
|
|
|
114 |
}
|
115 |
+
_atomic_write(STATE_PATH, state)
|
116 |
+
log.info(f"New Oracle prompt #{prompt_idx}: {state['prompt'][:80]}…")
|
117 |
return state
|
118 |
|
119 |
|
120 |
+
def _elapsed_str(start: float) -> str:
|
121 |
+
d = int(time.time() - start)
|
122 |
+
h, r = divmod(d, 3600)
|
123 |
+
m, s = divmod(r, 60)
|
124 |
+
return f"{h}h {m}m {s}s"
|
125 |
+
|
126 |
+
|
127 |
def oracle_loop():
|
|
|
128 |
while True:
|
129 |
with lock:
|
130 |
+
state = _init_state()
|
131 |
+
if state["finished"]:
|
132 |
+
time.sleep(SECS_BETWEEN_TOKENS)
|
133 |
+
continue
|
134 |
+
|
135 |
+
# Build context: prompt + generated so far
|
136 |
+
context = state["prompt"] + state["generated"]
|
137 |
+
input_ids = tokenizer(context, return_tensors="pt", truncation=True, max_length=MAX_CONTEXT_TOKENS).input_ids
|
138 |
+
|
139 |
+
# Sample one token
|
|
|
|
|
|
|
|
|
140 |
with torch.no_grad():
|
141 |
+
out = model.generate(
|
142 |
input_ids,
|
143 |
max_new_tokens=1,
|
144 |
do_sample=True,
|
145 |
temperature=TEMPERATURE,
|
146 |
top_p=TOP_P,
|
147 |
)
|
148 |
+
next_token = tokenizer.decode(out[0, -1], skip_special_tokens=True, clean_up_tokenization_spaces=False)
|
|
|
149 |
|
150 |
with lock:
|
151 |
+
state["generated"] += next_token
|
|
|
152 |
state["tokens_done"] += 1
|
153 |
if state["tokens_done"] >= TOKENS_PER_PROMPT:
|
154 |
state["finished"] = True
|
155 |
+
log.info("Prompt completed – Oracle will select a new one shortly.")
|
156 |
+
_atomic_write(STATE_PATH, state)
|
157 |
+
time.sleep(SECS_BETWEEN_TOKENS)
|
|
|
158 |
|
159 |
threading.Thread(target=oracle_loop, daemon=True).start()
|
160 |
|
161 |
+
###############################################################################
|
162 |
+
# Gradio interface #
|
163 |
+
###############################################################################
|
164 |
|
165 |
+
def fetch_state() -> tuple[str, str, str]:
|
166 |
+
state = _read_json(STATE_PATH, {})
|
167 |
+
if not state:
|
168 |
+
return "Loading…", "", "0h 0m 0s"
|
169 |
+
return state["prompt"], state["generated"], _elapsed_str(state["start_time"])
|
170 |
|
171 |
|
172 |
+
def submit_guess(full: str, idea: str):
|
173 |
+
full = full.strip()
|
174 |
+
idea = idea.strip()
|
175 |
+
if not full and not idea:
|
176 |
+
return gr.update(value="⚠️ Enter a guess in one of the fields."), gr.update(), gr.update()
|
|
|
|
|
|
|
|
|
|
|
177 |
|
178 |
+
prompt, generated, elapsed = fetch_state()
|
179 |
+
guess_text = full or idea
|
180 |
+
guess_type = "full" if full else "idea"
|
181 |
|
|
|
|
|
|
|
|
|
|
|
|
|
182 |
record = {
|
183 |
"timestamp": datetime.now(timezone.utc).isoformat(),
|
184 |
+
"prompt": prompt,
|
185 |
+
"point-in-time": elapsed,
|
186 |
+
"response-point": generated,
|
187 |
+
"user-guess": guess_text,
|
188 |
+
"guess-type": guess_type
|
189 |
}
|
|
|
190 |
with lock:
|
191 |
+
with open(DATA_PATH, "a", encoding="utf-8") as f:
|
192 |
f.write(json.dumps(record, ensure_ascii=False) + "\n")
|
193 |
+
log.info(f"Logged {guess_type} guess ({len(guess_text)} chars).")
|
194 |
+
return gr.update(value="✅ Guess recorded – thanks!"), gr.update(value=""), gr.update(value="")
|
195 |
|
196 |
|
197 |
+
with gr.Blocks(title="What Comes Next", theme="gradio/soft") as demo:
|
198 |
+
gr.Markdown("""# 🌌 What Comes Next
|
199 |
+
Watch the Oracle craft an extended response – **one token at a time**. Predict its
|
200 |
+
next words or general direction and see how close you were when the tale concludes.
|
201 |
+
(All inputs are stored in `data.json` for research.)""")
|
202 |
|
203 |
+
prompt_md = gr.Markdown()
|
204 |
+
oracle_box = gr.Textbox(lines=10, interactive=False, label="📜 Oracle text so far")
|
205 |
+
elapsed_tb = gr.Textbox(interactive=False, label="⏱ Elapsed time")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
206 |
|
207 |
+
refresh_btn = gr.Button("🔄 Refresh")
|
|
|
208 |
|
209 |
+
with gr.Row():
|
210 |
+
exact_tb = gr.Textbox(label="🧠 Exact continuation (full)")
|
211 |
+
idea_tb = gr.Textbox(label="💡 General idea")
|
212 |
+
submit_btn = gr.Button("Submit Guess")
|
213 |
+
status_tb = gr.Textbox(interactive=False, label="Status")
|
214 |
+
|
215 |
+
# Actions
|
216 |
+
refresh_btn.click(fetch_state, outputs=[prompt_md, oracle_box, elapsed_tb])
|
217 |
+
demo.load(fetch_state, outputs=[prompt_md, oracle_box, elapsed_tb])
|
218 |
+
|
219 |
+
submit_btn.click(submit_guess,
|
220 |
+
inputs=[exact_tb, idea_tb],
|
221 |
+
outputs=[status_tb, exact_tb, idea_tb])
|
222 |
|
223 |
if __name__ == "__main__":
|
224 |
demo.launch(server_name="0.0.0.0", server_port=7860, show_error=True)
|