Skip to content

feat(gdpval): raise per-call output and episode limits in sandbox recipes - #2

Closed
AdamRajfer wants to merge 1 commit into
gdpval-opencode-openclaw-sandbox-recipesfrom
gdpval-sandbox-recipe-limits
Closed

AdamRajfer wants to merge 1 commit into
gdpval-opencode-openclaw-sandbox-recipesfrom
gdpval-sandbox-recipe-limits

Conversation

@AdamRajfer

Copy link
Copy Markdown
Owner

Stacked on NVIDIA-NeMo#12. Recipe-level limits only; agent defaults are unchanged.

  • gdpval_openclaw_sandbox.yaml: max_output_tokens: 32768, the same per-call output limit the OpenCode recipe uses. Needs the native-mode change in fix(openclaw): honor max_output_tokens in sandbox sessions yaoyu-33/Gym#13 (otherwise OpenClaw rejects the field at activation).
  • gdpval_opencode_sandbox.yaml: timeout: 7200 per episode; with the previous limit a tenth of the tasks timed out before producing a deliverable.

Draft until the rerun of both recipes with these values is compared against the runs without them.

🤖 Generated with Claude Code

…ipes

OpenClaw: request 32768 output tokens per model call, matching the
OpenCode recipe, instead of the harness default. OpenCode: allow 7200 s
per episode. Recipe-level only; agent defaults are unchanged.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Signed-off-by: Adam Rajfer <arajfer@nvidia.com>
@AdamRajfer

Copy link
Copy Markdown
Owner Author

Moved to yaoyu-33#14 so it sits on the same stack as NVIDIA-NeMo#12.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant