NousResearch/atropos
1
0
Fork 0
mirror of https://github.com/NousResearch/atropos.git synced 2026-04-19 12:57:58 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

fix: handle identical scores in process mode with noise instead of None

Browse source
This commit is contained in:
RUFFY-369 2026-03-24 15:55:33 +05:30
parent 5c2afa8ea7
commit 9e727ce5ca
1 changed files with 29 additions and 7 deletions
Download patch file Download diff file Expand all files Collapse all files

36
environments/community/code_debug_env/code_debug_env.py
View file

@ -398,12 +398,24 @@ class CodeDebugEnv(BaseEnv):
reward, is_partial = self._score_fix(generated_code, item)
self.partial_fix_buffer.append(1 if is_partial else 0)
# Debug logging
extracted_preview = (
generated_code[:80] + "..." if generated_code else "(none)"
)
print(
f" [{item['entry_point']}] extracted={bool(generated_code)}, "
f"score={reward:.2f}, partial={is_partial}, "
f"code_preview={extracted_preview}"
)
tokens = rollout["tokens"]
masks = rollout["masks"]
logprobs = rollout["logprobs"]
# Remove obviously bad examples (too short responses)
if len([1 for m in masks if m != -100]) < 10:
non_masked = len([1 for m in masks if m != -100])
if non_masked < 10:
print(f" Skipping: only {non_masked} non-masked tokens")
continue
scores["tokens"].append(tokens)
@ -414,12 +426,16 @@ class CodeDebugEnv(BaseEnv):
if len(scores["tokens"]) >= self.config.group_size:
break
if not scores["scores"]:
print(f" WARNING: No valid rollouts for {item['entry_point']}")
return None
for score in scores["scores"]:
self.percent_correct_buffer.append(1.0 if score >= 1.0 else 0.0)
self.raw_score_buffer.append(max(score, 0.0))
# Apply length penalty when all scores are perfect
if scores["scores"] and all(score == 1.0 for score in scores["scores"]):
if all(score == 1.0 for score in scores["scores"]):
token_lengths = [len(t) for t in scores["tokens"]]
if max(token_lengths) == 0:
return None
@ -438,11 +454,17 @@ class CodeDebugEnv(BaseEnv):
pct = min(pct, 1.0)
scores["scores"].append(1.0 - pct)
# If all scores are same, return None (GRPO needs variance)
if scores["scores"] and all(
scores["scores"][0] == s for s in scores["scores"]
):
return None
# If all scores are same, add small noise to break ties
# This prevents infinite retry loops in process mode while
# maintaining meaningful relative ordering for training
if all(scores["scores"][0] == s for s in scores["scores"]):
print(
f" All {len(scores['scores'])} scores identical "
f"({scores['scores'][0]:.2f}), adding noise"
)
scores["scores"] = [
s + random.uniform(-0.01, 0.01) for s in scores["scores"]
]
return scores