maxent_grpo.training.seed_paper_eval_callback

Trainer callback for official SEED paper-style eval against the live vLLM server.

Functions

_coordination_dir(training_args, *, step)

_define_metric_axis(run, *, prefix)

_env_int(keys)

_latest_summary_path(results_dir)

_log_step0_summary_to_wandb(*, summary, ...)

_log_summary_to_wandb(*, summary, ...)

_mark_rank_release(coord_dir, *, rank, ...)

_repo_root()

_resolve_model_name(training_args)

_resolve_model_name_from_recipe_env()

_resolve_process_rank(training_args)

_resolve_python(training_args)

_resolve_results_root(training_args)

_resolve_step0_results_dir(training_args)

_resolve_vllm_url()

_resolve_workspace_dir(training_args)

_resolve_world_size(training_args)

_slugify(raw)

_sync_step0_summary_to_current_run(training_args)

_wait_for_rank_arrivals(coord_dir, *, ...)

_wait_for_rank_releases(coord_dir, *, ...)

_wait_for_result_payload(result_path, *, ...)

_wandb_run()

_write_json_atomic(path, payload)

build_live_seed_paper_eval_command(...)

Classes

SeedPaperEvalCallback(training_args)

Run the official SEED paper-style eval against the live vLLM server.
class maxent_grpo.training.seed_paper_eval_callback.SeedPaperEvalCallback(training_args)[source]

Bases: <class ‘TrainerCallback’>

Run the official SEED paper-style eval against the live vLLM server.

Parameters:

training_args (Any)

on_train_begin(args, state, control, **kwargs)[source]
Parameters:
Return type:

Any

on_step_end(args, state, control, **kwargs)[source]
Parameters:
Return type:

Any

on_evaluate(args, state, control, **kwargs)[source]
Parameters:
Return type:

Any

on_train_end(args, state, control, **kwargs)[source]
Parameters:
Return type:

Any

maxent_grpo.training.seed_paper_eval_callback.build_live_seed_paper_eval_command(training_args, *, step)[source]
Parameters:

  • training_args (Any)

  • step (int)

Return type:

tuple[list[str], Path]