Allow resuming vf-eval

[mikasenghaas]

Description

This PR adds support for resuming an evaluation run via vf-eval using the -R (--resume-from-path) flag which is supposed to be used in conjunction with -s and -f to allow to resume an evaluation run. It will:

• Load all completed example_ids from results.jsonl in the specified resume path
• Skip the already finished rollouts from the eval inputs
• Append new rollout results to results.jsonl and update the metadata (e.g. timings are summed, avg. reward and metrics are averaged)

Notes:

• This is a not a "full" checkpointing since we do not recover mid-trajectory failures but its the simplest solution to quickly resume a . Useful in long-running evals or synthetic data gen.
• Intermediate saving is only supported on the group level. If one needs rollout-level intermediate saving, it makes sense to duplicate prompts before running vf-eval and run with rollouts_per_example=1

Examples

Default behavior is unchanged

uv run vf-eval math500 -m gpt-5-nano -s

To use the resume feature and save each finished group, specify -s and -f 1. To illustrate, I sys.exit after every intermediate saving. For the initial command I run:

uv run vf-eval math500 -m gpt-5-nano -n 3 -r 1 -s -f 1

Then, each subsequent resume from the output dir

uv run vf-eval math500 -m gpt-5-nano -n 3 -r 1 -s -f 1 -R outputs/evals/math500--gpt-5-nano/67681b1d

Type of Change

• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality to not work as expected)
• Documentation update
• Test improvement

Testing

• All existing tests pass when running uv run pytest locally.
• New tests have been added to cover the changes

Checklist

• My code follows the style guidelines of this project as outlined in AGENTS.md
• I have performed a self-review of my own code
• I have commented my code, particularly in hard-to-understand areas
• I have made corresponding changes to the documentation
• My changes generate no new warnings
• Any dependent changes have been merged and published

Additional Notes