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

Merge pull request #408 from daspartho/verl-integration-fixes

Browse source 
fix: re-append stop string in math training path
This commit is contained in:
dmahan93 2026-03-10 23:08:58 -05:00 • committed by GitHub
commit c421582b6f
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
1 changed files with 4 additions and 0 deletions
Download patch file Download diff file Expand all files Collapse all files

4
environments/math_server_zero.py
View file

@ -472,6 +472,10 @@ class MathEnv(BaseEnv):
if self.config.mask_too_long_completions: if self.config.mask_too_long_completions:
scores["overrides"][-1]["set_advantage_to_zero"] = True scores["overrides"][-1]["set_advantage_to_zero"] = True
else: else:
# re-append </answer> if stripped by vLLM stop string handling
# (mirrors the eval path in rollout_and_score_eval)
if ("<answer>" in resp) and ("</answer>" not in resp):
resp = resp + "</answer>"
task = loop.run_in_executor(self.mp_executor, score_answer, gold, resp) task = loop.run_in_executor(self.mp_executor, score_answer, gold, resp)
reward = await task reward = await task
if reward is None: if reward is None: