Skip to content

disinguish is_truncated and log stop_conditions #2009

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
rasdani merged 1 commit into from
Mar 10, 2026
Merged

disinguish is_truncated and log stop_conditions #2009

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
9 changes: 9 additions & 0 deletions src/prime_rl/orchestrator/orchestrator.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -602,6 +602,7 @@ def process_rollout(rollout: vf.RolloutOutput, rollout_idx: int) -> list[Trainin
"reward": [rollout["reward"] for rollout in train_rollouts],
"is_truncated": [rollout["is_truncated"] for rollout in train_rollouts],
"error": [rollout["error"] for rollout in train_rollouts],
"stop_condition": [rollout.get("stop_condition") for rollout in train_rollouts],
"seq_len": [get_seq_len(rollout) for rollout in train_rollouts],
"prefill_len": rollout_prefill_lens,
"decode_len": rollout_decode_lens,
Expand Down Expand Up @@ -666,6 +667,14 @@ def process_rollout(rollout: vf.RolloutOutput, rollout_idx: int) -> list[Trainin
"is_truncated/mean": results_df.groupby("example_id").is_truncated.mean().mean(),
"is_truncated/max": results_df.groupby("example_id").is_truncated.mean().max(),
"is_truncated/min": results_df.groupby("example_id").is_truncated.mean().min(),
"stop_condition/generation_truncated": (
results_df.is_truncated & (results_df.stop_condition != "prompt_too_long")
).mean(),
Comment thread
rasdani marked this conversation as resolved.
# Log rate of each stop condition (e.g. max_turns, prompt_too_long, has_error)
**{
f"stop_condition/{sc}": rate
for sc, rate in results_df.stop_condition.dropna().value_counts(normalize=True).items()
},
# Seqs per rollout metrics
"samples_per_rollout/mean": results_df.groupby("example_id").samples_per_rollout.mean().mean(),
"samples_per_rollout/max": results_df.groupby("example_id").samples_per_rollout.mean().max(),
Expand Down