renpas22 commited on
Commit ·
464ac9b
1
Parent(s): 3024a91
Fix SPECIAL_TOKENS usage - import at module level and use string literals
Browse files
src/reasoning/step_level_cot.py
CHANGED
|
@@ -17,7 +17,7 @@ from diffusers import QwenImagePipeline, QwenImageTransformer2DModel
|
|
| 17 |
from omegaconf import OmegaConf
|
| 18 |
from tqdm.auto import tqdm
|
| 19 |
|
| 20 |
-
from .step_data import StepDataset, ReasoningChain
|
| 21 |
from .prm import ProcessRewardModel, PRMTrainer
|
| 22 |
from .rl_trainer import RLReasoningTrainer, RLConfig
|
| 23 |
from .inference_scaling import InferenceTimeScaling, InferenceConfig
|
|
@@ -390,10 +390,10 @@ class StepLevelCoTTrainer:
|
|
| 390 |
for chain in batch:
|
| 391 |
# Format: question + reasoning steps + answer
|
| 392 |
text = f"Question: {chain.prompt}\n"
|
| 393 |
-
text +=
|
| 394 |
for step in chain.steps:
|
| 395 |
-
text += f"
|
| 396 |
-
text +=
|
| 397 |
text += f"Answer: {chain.final_answer}"
|
| 398 |
|
| 399 |
texts.append(text)
|
|
|
|
| 17 |
from omegaconf import OmegaConf
|
| 18 |
from tqdm.auto import tqdm
|
| 19 |
|
| 20 |
+
from .step_data import StepDataset, ReasoningChain, SPECIAL_TOKENS
|
| 21 |
from .prm import ProcessRewardModel, PRMTrainer
|
| 22 |
from .rl_trainer import RLReasoningTrainer, RLConfig
|
| 23 |
from .inference_scaling import InferenceTimeScaling, InferenceConfig
|
|
|
|
| 390 |
for chain in batch:
|
| 391 |
# Format: question + reasoning steps + answer
|
| 392 |
text = f"Question: {chain.prompt}\n"
|
| 393 |
+
text += "<|reasoning_start|>\n"
|
| 394 |
for step in chain.steps:
|
| 395 |
+
text += f"<|step_start|>{step.content}<|step_end|>\n"
|
| 396 |
+
text += "<|reasoning_end|>\n"
|
| 397 |
text += f"Answer: {chain.final_answer}"
|
| 398 |
|
| 399 |
texts.append(text)
|