Commit
·
dbebf53
1
Parent(s):
e0e167f
Updated Hyperparams and dataset
Browse files- Gemma2_2B/finetune.ipynb +147 -345
- Gemma2_2B/hyperparams.yaml +13 -7
Gemma2_2B/finetune.ipynb
CHANGED
@@ -2,7 +2,7 @@
|
|
2 |
"cells": [
|
3 |
{
|
4 |
"cell_type": "code",
|
5 |
-
"execution_count":
|
6 |
"metadata": {},
|
7 |
"outputs": [],
|
8 |
"source": [
|
@@ -15,87 +15,84 @@
|
|
15 |
"login(token=os.getenv(\"HUGGINGFACE_TOKEN\"))"
|
16 |
]
|
17 |
},
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
18 |
{
|
19 |
"cell_type": "code",
|
20 |
-
"execution_count":
|
21 |
"metadata": {},
|
22 |
-
"outputs": [
|
23 |
-
{
|
24 |
-
"data": {
|
25 |
-
"application/vnd.jupyter.widget-view+json": {
|
26 |
-
"model_id": "a39e6120cbea4462999cfa5f887a8015",
|
27 |
-
"version_major": 2,
|
28 |
-
"version_minor": 0
|
29 |
-
},
|
30 |
-
"text/plain": [
|
31 |
-
"README.md: 0%| | 0.00/288 [00:00<?, ?B/s]"
|
32 |
-
]
|
33 |
-
},
|
34 |
-
"metadata": {},
|
35 |
-
"output_type": "display_data"
|
36 |
-
},
|
37 |
-
{
|
38 |
-
"name": "stderr",
|
39 |
-
"output_type": "stream",
|
40 |
-
"text": [
|
41 |
-
"f:\\TADBot\\.venv\\Lib\\site-packages\\huggingface_hub\\file_download.py:139: UserWarning: `huggingface_hub` cache-system uses symlinks by default to efficiently store duplicated files but your machine does not support them in C:\\Users\\Nitin Kausik Remella\\.cache\\huggingface\\hub\\datasets--ai-bites--databricks-mini. Caching files will still work but in a degraded version that might require more space on your disk. This warning can be disabled by setting the `HF_HUB_DISABLE_SYMLINKS_WARNING` environment variable. For more details, see https://huggingface.co/docs/huggingface_hub/how-to-cache#limitations.\n",
|
42 |
-
"To support symlinks on Windows, you either need to activate Developer Mode or to run Python as an administrator. In order to activate developer mode, see this article: https://docs.microsoft.com/en-us/windows/apps/get-started/enable-your-device-for-development\n",
|
43 |
-
" warnings.warn(message)\n"
|
44 |
-
]
|
45 |
-
},
|
46 |
-
{
|
47 |
-
"data": {
|
48 |
-
"application/vnd.jupyter.widget-view+json": {
|
49 |
-
"model_id": "de15e48751c34c36b5d02c2449380d06",
|
50 |
-
"version_major": 2,
|
51 |
-
"version_minor": 0
|
52 |
-
},
|
53 |
-
"text/plain": [
|
54 |
-
"dolly-mini-train.jsonl: 0%| | 0.00/5.24M [00:00<?, ?B/s]"
|
55 |
-
]
|
56 |
-
},
|
57 |
-
"metadata": {},
|
58 |
-
"output_type": "display_data"
|
59 |
-
},
|
60 |
-
{
|
61 |
-
"data": {
|
62 |
-
"application/vnd.jupyter.widget-view+json": {
|
63 |
-
"model_id": "d4094fd4af084a77a5bc3904b5db4197",
|
64 |
-
"version_major": 2,
|
65 |
-
"version_minor": 0
|
66 |
-
},
|
67 |
-
"text/plain": [
|
68 |
-
"Generating train split: 0%| | 0/10544 [00:00<?, ? examples/s]"
|
69 |
-
]
|
70 |
-
},
|
71 |
-
"metadata": {},
|
72 |
-
"output_type": "display_data"
|
73 |
-
},
|
74 |
-
{
|
75 |
-
"data": {
|
76 |
-
"text/plain": [
|
77 |
-
"Dataset({\n",
|
78 |
-
" features: ['text'],\n",
|
79 |
-
" num_rows: 1000\n",
|
80 |
-
"})"
|
81 |
-
]
|
82 |
-
},
|
83 |
-
"execution_count": 10,
|
84 |
-
"metadata": {},
|
85 |
-
"output_type": "execute_result"
|
86 |
-
}
|
87 |
-
],
|
88 |
"source": [
|
89 |
"from datasets import load_dataset\n",
|
90 |
-
"dataset_name = \"
|
91 |
-
"dataset = load_dataset(dataset_name, split=\"train
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
92 |
"\n",
|
93 |
-
"
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
94 |
]
|
95 |
},
|
96 |
{
|
97 |
"cell_type": "code",
|
98 |
-
"execution_count":
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
99 |
"metadata": {},
|
100 |
"outputs": [],
|
101 |
"source": [
|
@@ -109,12 +106,12 @@
|
|
109 |
" logging,\n",
|
110 |
")\n",
|
111 |
"from peft import LoraConfig, PeftModel\n",
|
112 |
-
"from trl import SFTTrainer"
|
113 |
]
|
114 |
},
|
115 |
{
|
116 |
"cell_type": "code",
|
117 |
-
"execution_count":
|
118 |
"metadata": {},
|
119 |
"outputs": [],
|
120 |
"source": [
|
@@ -125,7 +122,7 @@
|
|
125 |
},
|
126 |
{
|
127 |
"cell_type": "code",
|
128 |
-
"execution_count":
|
129 |
"metadata": {},
|
130 |
"outputs": [],
|
131 |
"source": [
|
@@ -141,17 +138,9 @@
|
|
141 |
},
|
142 |
{
|
143 |
"cell_type": "code",
|
144 |
-
"execution_count":
|
145 |
"metadata": {},
|
146 |
-
"outputs": [
|
147 |
-
{
|
148 |
-
"name": "stdout",
|
149 |
-
"output_type": "stream",
|
150 |
-
"text": [
|
151 |
-
"Setting BF16 to True\n"
|
152 |
-
]
|
153 |
-
}
|
154 |
-
],
|
155 |
"source": [
|
156 |
"# Check GPU compatibility with bfloat16\n",
|
157 |
"if compute_dtype == torch.float16 and hyperparams['use_4bit']:\n",
|
@@ -165,24 +154,9 @@
|
|
165 |
},
|
166 |
{
|
167 |
"cell_type": "code",
|
168 |
-
"execution_count":
|
169 |
"metadata": {},
|
170 |
-
"outputs": [
|
171 |
-
{
|
172 |
-
"data": {
|
173 |
-
"application/vnd.jupyter.widget-view+json": {
|
174 |
-
"model_id": "9ab84ef6c43249de9726940a78f2717f",
|
175 |
-
"version_major": 2,
|
176 |
-
"version_minor": 0
|
177 |
-
},
|
178 |
-
"text/plain": [
|
179 |
-
"Loading checkpoint shards: 0%| | 0/2 [00:00<?, ?it/s]"
|
180 |
-
]
|
181 |
-
},
|
182 |
-
"metadata": {},
|
183 |
-
"output_type": "display_data"
|
184 |
-
}
|
185 |
-
],
|
186 |
"source": [
|
187 |
"model = AutoModelForCausalLM.from_pretrained(\n",
|
188 |
" hyperparams['model_name'],\n",
|
@@ -201,7 +175,7 @@
|
|
201 |
},
|
202 |
{
|
203 |
"cell_type": "code",
|
204 |
-
"execution_count":
|
205 |
"metadata": {},
|
206 |
"outputs": [],
|
207 |
"source": [
|
@@ -218,275 +192,103 @@
|
|
218 |
},
|
219 |
{
|
220 |
"cell_type": "code",
|
221 |
-
"execution_count":
|
222 |
"metadata": {},
|
223 |
-
"outputs": [
|
224 |
-
{
|
225 |
-
"data": {
|
226 |
-
"text/plain": [
|
227 |
-
"TrainingArguments(\n",
|
228 |
-
"_n_gpu=1,\n",
|
229 |
-
"accelerator_config={'split_batches': False, 'dispatch_batches': None, 'even_batches': True, 'use_seedable_sampler': True, 'non_blocking': False, 'gradient_accumulation_kwargs': None, 'use_configured_state': False},\n",
|
230 |
-
"adafactor=False,\n",
|
231 |
-
"adam_beta1=0.9,\n",
|
232 |
-
"adam_beta2=0.999,\n",
|
233 |
-
"adam_epsilon=1e-08,\n",
|
234 |
-
"auto_find_batch_size=False,\n",
|
235 |
-
"average_tokens_across_devices=False,\n",
|
236 |
-
"batch_eval_metrics=False,\n",
|
237 |
-
"bf16=True,\n",
|
238 |
-
"bf16_full_eval=False,\n",
|
239 |
-
"data_seed=None,\n",
|
240 |
-
"dataloader_drop_last=False,\n",
|
241 |
-
"dataloader_num_workers=0,\n",
|
242 |
-
"dataloader_persistent_workers=False,\n",
|
243 |
-
"dataloader_pin_memory=True,\n",
|
244 |
-
"dataloader_prefetch_factor=None,\n",
|
245 |
-
"ddp_backend=None,\n",
|
246 |
-
"ddp_broadcast_buffers=None,\n",
|
247 |
-
"ddp_bucket_cap_mb=None,\n",
|
248 |
-
"ddp_find_unused_parameters=None,\n",
|
249 |
-
"ddp_timeout=1800,\n",
|
250 |
-
"debug=[],\n",
|
251 |
-
"deepspeed=None,\n",
|
252 |
-
"disable_tqdm=False,\n",
|
253 |
-
"dispatch_batches=None,\n",
|
254 |
-
"do_eval=False,\n",
|
255 |
-
"do_predict=False,\n",
|
256 |
-
"do_train=False,\n",
|
257 |
-
"eval_accumulation_steps=None,\n",
|
258 |
-
"eval_delay=0,\n",
|
259 |
-
"eval_do_concat_batches=True,\n",
|
260 |
-
"eval_on_start=False,\n",
|
261 |
-
"eval_steps=None,\n",
|
262 |
-
"eval_strategy=IntervalStrategy.NO,\n",
|
263 |
-
"eval_use_gather_object=False,\n",
|
264 |
-
"evaluation_strategy=None,\n",
|
265 |
-
"fp16=False,\n",
|
266 |
-
"fp16_backend=auto,\n",
|
267 |
-
"fp16_full_eval=False,\n",
|
268 |
-
"fp16_opt_level=O1,\n",
|
269 |
-
"fsdp=[],\n",
|
270 |
-
"fsdp_config={'min_num_params': 0, 'xla': False, 'xla_fsdp_v2': False, 'xla_fsdp_grad_ckpt': False},\n",
|
271 |
-
"fsdp_min_num_params=0,\n",
|
272 |
-
"fsdp_transformer_layer_cls_to_wrap=None,\n",
|
273 |
-
"full_determinism=False,\n",
|
274 |
-
"gradient_accumulation_steps=1,\n",
|
275 |
-
"gradient_checkpointing=False,\n",
|
276 |
-
"gradient_checkpointing_kwargs=None,\n",
|
277 |
-
"greater_is_better=None,\n",
|
278 |
-
"group_by_length=True,\n",
|
279 |
-
"half_precision_backend=auto,\n",
|
280 |
-
"hub_always_push=False,\n",
|
281 |
-
"hub_model_id=None,\n",
|
282 |
-
"hub_private_repo=False,\n",
|
283 |
-
"hub_strategy=HubStrategy.EVERY_SAVE,\n",
|
284 |
-
"hub_token=<HUB_TOKEN>,\n",
|
285 |
-
"ignore_data_skip=False,\n",
|
286 |
-
"include_for_metrics=[],\n",
|
287 |
-
"include_inputs_for_metrics=False,\n",
|
288 |
-
"include_num_input_tokens_seen=False,\n",
|
289 |
-
"include_tokens_per_second=False,\n",
|
290 |
-
"jit_mode_eval=False,\n",
|
291 |
-
"label_names=None,\n",
|
292 |
-
"label_smoothing_factor=0.0,\n",
|
293 |
-
"learning_rate=0.0002,\n",
|
294 |
-
"length_column_name=length,\n",
|
295 |
-
"load_best_model_at_end=False,\n",
|
296 |
-
"local_rank=0,\n",
|
297 |
-
"log_level=passive,\n",
|
298 |
-
"log_level_replica=warning,\n",
|
299 |
-
"log_on_each_node=True,\n",
|
300 |
-
"logging_dir=./results\\runs\\Nov15_13-14-10_FutureGadgetLab,\n",
|
301 |
-
"logging_first_step=False,\n",
|
302 |
-
"logging_nan_inf_filter=True,\n",
|
303 |
-
"logging_steps=25,\n",
|
304 |
-
"logging_strategy=IntervalStrategy.STEPS,\n",
|
305 |
-
"lr_scheduler_kwargs={},\n",
|
306 |
-
"lr_scheduler_type=SchedulerType.CONSTANT,\n",
|
307 |
-
"max_grad_norm=0.3,\n",
|
308 |
-
"max_steps=-1,\n",
|
309 |
-
"metric_for_best_model=None,\n",
|
310 |
-
"mp_parameters=,\n",
|
311 |
-
"neftune_noise_alpha=None,\n",
|
312 |
-
"no_cuda=False,\n",
|
313 |
-
"num_train_epochs=1,\n",
|
314 |
-
"optim=OptimizerNames.PAGED_ADAMW,\n",
|
315 |
-
"optim_args=None,\n",
|
316 |
-
"optim_target_modules=None,\n",
|
317 |
-
"output_dir=./results,\n",
|
318 |
-
"overwrite_output_dir=False,\n",
|
319 |
-
"past_index=-1,\n",
|
320 |
-
"per_device_eval_batch_size=8,\n",
|
321 |
-
"per_device_train_batch_size=2,\n",
|
322 |
-
"prediction_loss_only=False,\n",
|
323 |
-
"push_to_hub=False,\n",
|
324 |
-
"push_to_hub_model_id=None,\n",
|
325 |
-
"push_to_hub_organization=None,\n",
|
326 |
-
"push_to_hub_token=<PUSH_TO_HUB_TOKEN>,\n",
|
327 |
-
"ray_scope=last,\n",
|
328 |
-
"remove_unused_columns=True,\n",
|
329 |
-
"report_to=['tensorboard'],\n",
|
330 |
-
"restore_callback_states_from_checkpoint=False,\n",
|
331 |
-
"resume_from_checkpoint=None,\n",
|
332 |
-
"run_name=./results,\n",
|
333 |
-
"save_on_each_node=False,\n",
|
334 |
-
"save_only_model=False,\n",
|
335 |
-
"save_safetensors=True,\n",
|
336 |
-
"save_steps=25,\n",
|
337 |
-
"save_strategy=IntervalStrategy.STEPS,\n",
|
338 |
-
"save_total_limit=None,\n",
|
339 |
-
"seed=42,\n",
|
340 |
-
"skip_memory_metrics=True,\n",
|
341 |
-
"split_batches=None,\n",
|
342 |
-
"tf32=None,\n",
|
343 |
-
"torch_compile=False,\n",
|
344 |
-
"torch_compile_backend=None,\n",
|
345 |
-
"torch_compile_mode=None,\n",
|
346 |
-
"torch_empty_cache_steps=None,\n",
|
347 |
-
"torchdynamo=None,\n",
|
348 |
-
"tpu_metrics_debug=False,\n",
|
349 |
-
"tpu_num_cores=None,\n",
|
350 |
-
"use_cpu=False,\n",
|
351 |
-
"use_ipex=False,\n",
|
352 |
-
"use_legacy_prediction_loop=False,\n",
|
353 |
-
"use_liger_kernel=False,\n",
|
354 |
-
"use_mps_device=False,\n",
|
355 |
-
"warmup_ratio=0.03,\n",
|
356 |
-
"warmup_steps=0,\n",
|
357 |
-
"weight_decay=0.001,\n",
|
358 |
-
")"
|
359 |
-
]
|
360 |
-
},
|
361 |
-
"execution_count": 39,
|
362 |
-
"metadata": {},
|
363 |
-
"output_type": "execute_result"
|
364 |
-
}
|
365 |
-
],
|
366 |
"source": [
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
367 |
"# Set training parameters\n",
|
368 |
"training_arguments = TrainingArguments(\n",
|
369 |
-
" output_dir=
|
370 |
-
"
|
371 |
-
"
|
372 |
-
" gradient_accumulation_steps=hyperparams[
|
373 |
-
" optim=hyperparams[
|
374 |
-
"
|
375 |
-
"
|
376 |
-
"
|
377 |
-
"
|
378 |
-
"
|
379 |
-
"
|
380 |
-
"
|
381 |
-
"
|
382 |
-
"
|
383 |
-
"
|
384 |
-
"
|
385 |
-
"
|
|
|
|
|
|
|
|
|
|
|
|
|
386 |
")\n",
|
387 |
"training_arguments"
|
388 |
]
|
389 |
},
|
390 |
{
|
391 |
"cell_type": "code",
|
392 |
-
"execution_count":
|
393 |
"metadata": {},
|
394 |
-
"outputs": [
|
395 |
-
{
|
396 |
-
"name": "stderr",
|
397 |
-
"output_type": "stream",
|
398 |
-
"text": [
|
399 |
-
"f:\\TADBot\\.venv\\Lib\\site-packages\\huggingface_hub\\utils\\_deprecation.py:100: FutureWarning: Deprecated argument(s) used in '__init__': dataset_text_field, max_seq_length, packing. Will not be supported from version '0.13.0'.\n",
|
400 |
-
"\n",
|
401 |
-
"Deprecated positional argument(s) used in SFTTrainer, please use the SFTConfig to set these arguments instead.\n",
|
402 |
-
" warnings.warn(message, FutureWarning)\n",
|
403 |
-
"f:\\TADBot\\.venv\\Lib\\site-packages\\trl\\trainer\\sft_trainer.py:212: UserWarning: You passed a `packing` argument to the SFTTrainer, the value you passed will override the one in the `SFTConfig`.\n",
|
404 |
-
" warnings.warn(\n",
|
405 |
-
"f:\\TADBot\\.venv\\Lib\\site-packages\\trl\\trainer\\sft_trainer.py:300: UserWarning: You passed a `max_seq_length` argument to the SFTTrainer, the value you passed will override the one in the `SFTConfig`.\n",
|
406 |
-
" warnings.warn(\n",
|
407 |
-
"f:\\TADBot\\.venv\\Lib\\site-packages\\trl\\trainer\\sft_trainer.py:328: UserWarning: You passed a `dataset_text_field` argument to the SFTTrainer, the value you passed will override the one in the `SFTConfig`.\n",
|
408 |
-
" warnings.warn(\n"
|
409 |
-
]
|
410 |
-
}
|
411 |
-
],
|
412 |
"source": [
|
413 |
"trainer = SFTTrainer(\n",
|
414 |
" model=model,\n",
|
415 |
-
" train_dataset=dataset,\n",
|
|
|
416 |
" peft_config=peft_config,\n",
|
417 |
" dataset_text_field=\"text\",\n",
|
418 |
" # formatting_func=format_prompts_fn,\n",
|
419 |
-
" max_seq_length=hyperparams[
|
420 |
" tokenizer=tokenizer,\n",
|
421 |
" args=training_arguments,\n",
|
422 |
-
" packing=hyperparams[
|
423 |
")"
|
424 |
]
|
425 |
},
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
426 |
{
|
427 |
"cell_type": "code",
|
428 |
"execution_count": null,
|
429 |
"metadata": {},
|
430 |
-
"outputs": [
|
431 |
-
|
432 |
-
|
433 |
-
|
434 |
-
|
435 |
-
|
436 |
-
|
437 |
-
|
438 |
-
|
439 |
-
|
440 |
-
|
441 |
-
},
|
442 |
-
"metadata": {},
|
443 |
-
"output_type": "display_data"
|
444 |
-
},
|
445 |
-
{
|
446 |
-
"name": "stdout",
|
447 |
-
"output_type": "stream",
|
448 |
-
"text": [
|
449 |
-
"{'loss': 3.8879, 'grad_norm': 18.030195236206055, 'learning_rate': 0.0002, 'epoch': 0.02}\n",
|
450 |
-
"{'loss': 2.9569, 'grad_norm': 9.667036056518555, 'learning_rate': 0.0002, 'epoch': 0.04}\n",
|
451 |
-
"{'loss': 2.6361, 'grad_norm': 9.089476585388184, 'learning_rate': 0.0002, 'epoch': 0.06}\n",
|
452 |
-
"{'loss': 2.9523, 'grad_norm': 6.053662300109863, 'learning_rate': 0.0002, 'epoch': 0.07}\n",
|
453 |
-
"{'loss': 2.8543, 'grad_norm': 7.764152526855469, 'learning_rate': 0.0002, 'epoch': 0.09}\n",
|
454 |
-
"{'loss': 2.8802, 'grad_norm': 6.539248466491699, 'learning_rate': 0.0002, 'epoch': 0.11}\n",
|
455 |
-
"{'loss': 2.7047, 'grad_norm': 5.485109329223633, 'learning_rate': 0.0002, 'epoch': 0.13}\n",
|
456 |
-
"{'loss': 2.6576, 'grad_norm': 9.22624397277832, 'learning_rate': 0.0002, 'epoch': 0.15}\n",
|
457 |
-
"{'loss': 2.7756, 'grad_norm': 6.477100372314453, 'learning_rate': 0.0002, 'epoch': 0.17}\n",
|
458 |
-
"{'loss': 2.7012, 'grad_norm': 5.891603946685791, 'learning_rate': 0.0002, 'epoch': 0.19}\n",
|
459 |
-
"{'loss': 2.5026, 'grad_norm': 5.75968599319458, 'learning_rate': 0.0002, 'epoch': 0.21}\n",
|
460 |
-
"{'loss': 2.8085, 'grad_norm': 7.938610076904297, 'learning_rate': 0.0002, 'epoch': 0.22}\n",
|
461 |
-
"{'loss': 2.5286, 'grad_norm': 5.600504398345947, 'learning_rate': 0.0002, 'epoch': 0.24}\n",
|
462 |
-
"{'loss': 2.5495, 'grad_norm': 6.746212005615234, 'learning_rate': 0.0002, 'epoch': 0.26}\n",
|
463 |
-
"{'loss': 2.7405, 'grad_norm': 3.8923749923706055, 'learning_rate': 0.0002, 'epoch': 0.28}\n",
|
464 |
-
"{'loss': 2.5657, 'grad_norm': 5.949460506439209, 'learning_rate': 0.0002, 'epoch': 0.3}\n",
|
465 |
-
"{'loss': 2.6052, 'grad_norm': 5.733223915100098, 'learning_rate': 0.0002, 'epoch': 0.32}\n",
|
466 |
-
"{'loss': 2.673, 'grad_norm': 6.0587310791015625, 'learning_rate': 0.0002, 'epoch': 0.34}\n",
|
467 |
-
"{'loss': 2.4631, 'grad_norm': 4.734077453613281, 'learning_rate': 0.0002, 'epoch': 0.35}\n",
|
468 |
-
"{'loss': 2.7288, 'grad_norm': 6.7847700119018555, 'learning_rate': 0.0002, 'epoch': 0.37}\n",
|
469 |
-
"{'loss': 2.7797, 'grad_norm': 5.118943214416504, 'learning_rate': 0.0002, 'epoch': 0.39}\n",
|
470 |
-
"{'loss': 2.8644, 'grad_norm': 5.4167304039001465, 'learning_rate': 0.0002, 'epoch': 0.41}\n",
|
471 |
-
"{'loss': 2.5779, 'grad_norm': 6.73247766494751, 'learning_rate': 0.0002, 'epoch': 0.43}\n",
|
472 |
-
"{'loss': 2.6459, 'grad_norm': 4.644010066986084, 'learning_rate': 0.0002, 'epoch': 0.45}\n",
|
473 |
-
"{'loss': 2.5321, 'grad_norm': 6.347738265991211, 'learning_rate': 0.0002, 'epoch': 0.47}\n",
|
474 |
-
"{'loss': 2.6865, 'grad_norm': 5.185911655426025, 'learning_rate': 0.0002, 'epoch': 0.49}\n",
|
475 |
-
"{'loss': 2.4668, 'grad_norm': 5.355742454528809, 'learning_rate': 0.0002, 'epoch': 0.5}\n",
|
476 |
-
"{'loss': 2.8465, 'grad_norm': 5.4434380531311035, 'learning_rate': 0.0002, 'epoch': 0.52}\n",
|
477 |
-
"{'loss': 2.7376, 'grad_norm': 4.8459882736206055, 'learning_rate': 0.0002, 'epoch': 0.54}\n",
|
478 |
-
"{'loss': 2.5205, 'grad_norm': 5.886116981506348, 'learning_rate': 0.0002, 'epoch': 0.56}\n",
|
479 |
-
"{'loss': 2.7473, 'grad_norm': 4.946981906890869, 'learning_rate': 0.0002, 'epoch': 0.58}\n",
|
480 |
-
"{'loss': 2.6824, 'grad_norm': 6.349016189575195, 'learning_rate': 0.0002, 'epoch': 0.6}\n",
|
481 |
-
"{'loss': 2.6485, 'grad_norm': 5.024953365325928, 'learning_rate': 0.0002, 'epoch': 0.62}\n",
|
482 |
-
"{'loss': 2.7172, 'grad_norm': 5.583380222320557, 'learning_rate': 0.0002, 'epoch': 0.63}\n",
|
483 |
-
"{'loss': 2.5879, 'grad_norm': 6.582890033721924, 'learning_rate': 0.0002, 'epoch': 0.65}\n"
|
484 |
-
]
|
485 |
-
}
|
486 |
-
],
|
487 |
"source": [
|
488 |
-
"
|
489 |
-
"trainer.model.save_pretrained(hyperparams['new_model_name'])"
|
490 |
]
|
491 |
}
|
492 |
],
|
|
|
2 |
"cells": [
|
3 |
{
|
4 |
"cell_type": "code",
|
5 |
+
"execution_count": null,
|
6 |
"metadata": {},
|
7 |
"outputs": [],
|
8 |
"source": [
|
|
|
15 |
"login(token=os.getenv(\"HUGGINGFACE_TOKEN\"))"
|
16 |
]
|
17 |
},
|
18 |
+
{
|
19 |
+
"cell_type": "markdown",
|
20 |
+
"metadata": {},
|
21 |
+
"source": [
|
22 |
+
"# Dataset\n",
|
23 |
+
"Modifyify the dataset to fit the Gemma 2 prompt format"
|
24 |
+
]
|
25 |
+
},
|
26 |
{
|
27 |
"cell_type": "code",
|
28 |
+
"execution_count": null,
|
29 |
"metadata": {},
|
30 |
+
"outputs": [],
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
31 |
"source": [
|
32 |
"from datasets import load_dataset\n",
|
33 |
+
"dataset_name = \"nbertagnolli/counsel-chat\"\n",
|
34 |
+
"dataset = load_dataset(dataset_name, split=\"train\",cache_dir=\".cache/\")\n",
|
35 |
+
"\n",
|
36 |
+
"# Print the first example from the dataset\n",
|
37 |
+
"print(dataset[0])\n",
|
38 |
+
"print(f\"\\n {dataset}\")"
|
39 |
+
]
|
40 |
+
},
|
41 |
+
{
|
42 |
+
"cell_type": "code",
|
43 |
+
"execution_count": null,
|
44 |
+
"metadata": {},
|
45 |
+
"outputs": [],
|
46 |
+
"source": [
|
47 |
+
"gemma_prompt = \"\"\" \n",
|
48 |
+
"### System:\n",
|
49 |
+
"You are a Therapist Assistant, an LLM fine-tuned on Gemma 2 model by Google.\n",
|
50 |
+
"You provide safe and responsible support to users while encouraging them to visit a mental health professional if needed. \n",
|
51 |
+
"You are committed to promoting wellness, understanding, and support. Your responses should be clear, concise, and evidence-based, while maintaining a friendly and approachable tone.\n",
|
52 |
"\n",
|
53 |
+
"### User:\n",
|
54 |
+
"{}\n",
|
55 |
+
"\n",
|
56 |
+
"### Response:\n",
|
57 |
+
"{}\n",
|
58 |
+
"\"\"\"\n",
|
59 |
+
"\n",
|
60 |
+
"def format_prompts_func(example):\n",
|
61 |
+
" \"\"\"Formats questionText and answerText into the Gemma 2 prompt format.\"\"\"\n",
|
62 |
+
" question_texts = example[\"questionText\"]\n",
|
63 |
+
" answer_texts = example[\"answerText\"]\n",
|
64 |
+
" texts = []\n",
|
65 |
+
" for q, a in zip(question_texts, answer_texts):\n",
|
66 |
+
" text = gemma_prompt.format(q, a)\n",
|
67 |
+
" texts.append(text)\n",
|
68 |
+
"\n",
|
69 |
+
" return {\"text\": texts}\n",
|
70 |
+
"pass\n",
|
71 |
+
"# Apply the formatting function to the dataset\n",
|
72 |
+
"formatted_dataset = dataset.map(format_prompts_func, batched=True)\n",
|
73 |
+
"print(formatted_dataset['text'][0])\n"
|
74 |
]
|
75 |
},
|
76 |
{
|
77 |
"cell_type": "code",
|
78 |
+
"execution_count": null,
|
79 |
+
"metadata": {},
|
80 |
+
"outputs": [],
|
81 |
+
"source": [
|
82 |
+
"dataset = formatted_dataset.train_test_split(test_size=0.2, seed=42)\n",
|
83 |
+
"print(dataset['train'].shape, dataset['test'].shape)"
|
84 |
+
]
|
85 |
+
},
|
86 |
+
{
|
87 |
+
"cell_type": "markdown",
|
88 |
+
"metadata": {},
|
89 |
+
"source": [
|
90 |
+
"# Fine tuning hyperpterparameters"
|
91 |
+
]
|
92 |
+
},
|
93 |
+
{
|
94 |
+
"cell_type": "code",
|
95 |
+
"execution_count": null,
|
96 |
"metadata": {},
|
97 |
"outputs": [],
|
98 |
"source": [
|
|
|
106 |
" logging,\n",
|
107 |
")\n",
|
108 |
"from peft import LoraConfig, PeftModel\n",
|
109 |
+
"from trl import SFTTrainer\n"
|
110 |
]
|
111 |
},
|
112 |
{
|
113 |
"cell_type": "code",
|
114 |
+
"execution_count": null,
|
115 |
"metadata": {},
|
116 |
"outputs": [],
|
117 |
"source": [
|
|
|
122 |
},
|
123 |
{
|
124 |
"cell_type": "code",
|
125 |
+
"execution_count": null,
|
126 |
"metadata": {},
|
127 |
"outputs": [],
|
128 |
"source": [
|
|
|
138 |
},
|
139 |
{
|
140 |
"cell_type": "code",
|
141 |
+
"execution_count": null,
|
142 |
"metadata": {},
|
143 |
+
"outputs": [],
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
144 |
"source": [
|
145 |
"# Check GPU compatibility with bfloat16\n",
|
146 |
"if compute_dtype == torch.float16 and hyperparams['use_4bit']:\n",
|
|
|
154 |
},
|
155 |
{
|
156 |
"cell_type": "code",
|
157 |
+
"execution_count": null,
|
158 |
"metadata": {},
|
159 |
+
"outputs": [],
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
160 |
"source": [
|
161 |
"model = AutoModelForCausalLM.from_pretrained(\n",
|
162 |
" hyperparams['model_name'],\n",
|
|
|
175 |
},
|
176 |
{
|
177 |
"cell_type": "code",
|
178 |
+
"execution_count": null,
|
179 |
"metadata": {},
|
180 |
"outputs": [],
|
181 |
"source": [
|
|
|
192 |
},
|
193 |
{
|
194 |
"cell_type": "code",
|
195 |
+
"execution_count": null,
|
196 |
"metadata": {},
|
197 |
+
"outputs": [],
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
198 |
"source": [
|
199 |
+
"import wandb\n",
|
200 |
+
"import time\n",
|
201 |
+
"wandb.login(key=os.getenv(\"WANDB_API_KEY\"))\n",
|
202 |
+
"run = wandb.init(\n",
|
203 |
+
" project='TADBot',\n",
|
204 |
+
" job_type=\"training\",\n",
|
205 |
+
" anonymous=\"allow\"\n",
|
206 |
+
")\n",
|
207 |
+
"run_name = f\"{hyperparams['model_name']}--health-bot-{int(time.time())}\"\n",
|
208 |
+
"\n",
|
209 |
"# Set training parameters\n",
|
210 |
"training_arguments = TrainingArguments(\n",
|
211 |
+
" output_dir=f\"./outputs/{run_name}\",\n",
|
212 |
+
" per_device_train_batch_size=hyperparams[\"per_device_train_batch_size\"],\n",
|
213 |
+
" per_device_eval_batch_size=hyperparams[\"per_device_eval_batch_size\"],\n",
|
214 |
+
" gradient_accumulation_steps=hyperparams[\"gradient_accumulation_steps\"],\n",
|
215 |
+
" optim=hyperparams[\"optimizer\"],\n",
|
216 |
+
" num_train_epochs=hyperparams[\"num_train_epochs\"],\n",
|
217 |
+
" eval_steps=hyperparams[\"eval_steps\"],\n",
|
218 |
+
" eval_strategy=hyperparams[\"eval_strategy\"],\n",
|
219 |
+
" save_steps=hyperparams[\"save_steps\"],\n",
|
220 |
+
" logging_steps=hyperparams[\"logging_steps\"],\n",
|
221 |
+
" logging_strategy=hyperparams[\"logging_strategy\"],\n",
|
222 |
+
" warmup_steps=hyperparams[\"warmup_steps\"],\n",
|
223 |
+
" learning_rate=float(hyperparams[\"learning_rate\"]),\n",
|
224 |
+
" weight_decay=hyperparams[\"weight_decay\"],\n",
|
225 |
+
" fp16=hyperparams[\"fp16\"],\n",
|
226 |
+
" bf16=hyperparams[\"bf16\"],\n",
|
227 |
+
" max_grad_norm=hyperparams[\"max_grad_norm\"],\n",
|
228 |
+
" max_steps=hyperparams[\"max_steps\"],\n",
|
229 |
+
" group_by_length=hyperparams[\"group_by_length\"],\n",
|
230 |
+
" lr_scheduler_type=hyperparams[\"lr_scheduler_type\"],\n",
|
231 |
+
" logging_dir=f\"./outputs/{run_name}/logs\",\n",
|
232 |
+
" report_to=\"wandb\",\n",
|
233 |
+
" run_name=run_name\n",
|
234 |
")\n",
|
235 |
"training_arguments"
|
236 |
]
|
237 |
},
|
238 |
{
|
239 |
"cell_type": "code",
|
240 |
+
"execution_count": null,
|
241 |
"metadata": {},
|
242 |
+
"outputs": [],
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
243 |
"source": [
|
244 |
"trainer = SFTTrainer(\n",
|
245 |
" model=model,\n",
|
246 |
+
" train_dataset=dataset[\"train\"],\n",
|
247 |
+
" eval_dataset=dataset['test'],\n",
|
248 |
" peft_config=peft_config,\n",
|
249 |
" dataset_text_field=\"text\",\n",
|
250 |
" # formatting_func=format_prompts_fn,\n",
|
251 |
+
" max_seq_length=hyperparams[\"max_seq_length\"],\n",
|
252 |
" tokenizer=tokenizer,\n",
|
253 |
" args=training_arguments,\n",
|
254 |
+
" packing=hyperparams[\"packing\"],\n",
|
255 |
")"
|
256 |
]
|
257 |
},
|
258 |
+
{
|
259 |
+
"cell_type": "markdown",
|
260 |
+
"metadata": {},
|
261 |
+
"source": [
|
262 |
+
"# Fine tuning the model"
|
263 |
+
]
|
264 |
+
},
|
265 |
+
{
|
266 |
+
"cell_type": "code",
|
267 |
+
"execution_count": null,
|
268 |
+
"metadata": {},
|
269 |
+
"outputs": [],
|
270 |
+
"source": [
|
271 |
+
"model.config.use_cache = False\n",
|
272 |
+
"trainer.train()"
|
273 |
+
]
|
274 |
+
},
|
275 |
{
|
276 |
"cell_type": "code",
|
277 |
"execution_count": null,
|
278 |
"metadata": {},
|
279 |
+
"outputs": [],
|
280 |
+
"source": [
|
281 |
+
"wandb.finish()\n",
|
282 |
+
"model.config.use_cache = True\n",
|
283 |
+
"# Save the model\n",
|
284 |
+
"trainer.model.save_pretrained(hyperparams[\"new_model_name\"])"
|
285 |
+
]
|
286 |
+
},
|
287 |
+
{
|
288 |
+
"cell_type": "markdown",
|
289 |
+
"metadata": {},
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
290 |
"source": [
|
291 |
+
"%tensorboard --logdir Gemma2_2B\\\\results\\\\runs"
|
|
|
292 |
]
|
293 |
}
|
294 |
],
|
Gemma2_2B/hyperparams.yaml
CHANGED
@@ -1,34 +1,40 @@
|
|
1 |
model_name: "google/gemma-2-2b-it"
|
2 |
new_model_name: "gemma-2-2b-ft"
|
3 |
|
|
|
4 |
lora_r: 64
|
5 |
lora_alpha: 16
|
6 |
lora_dropout: 0.1
|
7 |
|
|
|
8 |
use_4bit: True
|
9 |
bnb_4bit_compute_dtype: "float16"
|
10 |
bnb_4bit_quant_type: "nf4"
|
11 |
use_nested_quant: False
|
12 |
|
13 |
-
|
14 |
-
num_train_epochs:
|
15 |
fp16: False
|
16 |
bf16: False
|
17 |
per_device_train_batch_size: 2
|
18 |
per_device_eval_batch_size: 2
|
19 |
-
gradient_accumulation_steps:
|
20 |
gradient_checkpointing: True
|
|
|
|
|
21 |
max_grad_norm: 0.3
|
22 |
-
learning_rate: 2e-
|
23 |
weight_decay: 0.001
|
24 |
optimizer: "paged_adamw_32bit"
|
25 |
lr_scheduler_type: "constant"
|
26 |
max_steps: -1
|
27 |
-
|
28 |
group_by_length: True
|
29 |
-
save_steps:
|
30 |
-
logging_steps:
|
|
|
31 |
|
|
|
32 |
max_seq_length: 128
|
33 |
packing: True
|
34 |
device_map: "auto"
|
|
|
1 |
model_name: "google/gemma-2-2b-it"
|
2 |
new_model_name: "gemma-2-2b-ft"
|
3 |
|
4 |
+
# LoRA Paraments
|
5 |
lora_r: 64
|
6 |
lora_alpha: 16
|
7 |
lora_dropout: 0.1
|
8 |
|
9 |
+
#bitsandbytes parameters
|
10 |
use_4bit: True
|
11 |
bnb_4bit_compute_dtype: "float16"
|
12 |
bnb_4bit_quant_type: "nf4"
|
13 |
use_nested_quant: False
|
14 |
|
15 |
+
#Training Arguments
|
16 |
+
num_train_epochs: 1
|
17 |
fp16: False
|
18 |
bf16: False
|
19 |
per_device_train_batch_size: 2
|
20 |
per_device_eval_batch_size: 2
|
21 |
+
gradient_accumulation_steps: 2
|
22 |
gradient_checkpointing: True
|
23 |
+
eval_strategy: "steps"
|
24 |
+
eval_steps: 0.2
|
25 |
max_grad_norm: 0.3
|
26 |
+
learning_rate: 2e-4
|
27 |
weight_decay: 0.001
|
28 |
optimizer: "paged_adamw_32bit"
|
29 |
lr_scheduler_type: "constant"
|
30 |
max_steps: -1
|
31 |
+
warmup_steps: 5
|
32 |
group_by_length: True
|
33 |
+
save_steps: 50
|
34 |
+
logging_steps: 50
|
35 |
+
logging_strategy: "steps"
|
36 |
|
37 |
+
#SFT Arguments
|
38 |
max_seq_length: 128
|
39 |
packing: True
|
40 |
device_map: "auto"
|