sheng22213/pku-safe-rlhf-125-eval-bundle

PKU SafeRLHF 125-Subset Evaluation Bundle

This bundle contains the code and prepared data used for the 125-sample evaluation subset under:

• —figure6_eval728_without_train_overlap_125_repro_seed0

It is intended as a portable reproduction package for:

• —generation on the fixed 125-prompt subset
• —pairwise helpfulness / harmlessness judging
• —single-model humanness judging
• —the newer mark-error humanness judging

Layout

• —code/: generation, judging, merge, and cluster helper scripts
• —data/: prepared 125-prompt subset and base-model generations
• —prompts/: prompt template used by the mark-error humanness judge
• —metadata/: manifest and source-address notes

Included Data

• —data/evaluation_prompts-full-728_without_train_overlap_125_random_seed0.json
• —data/evaluation_prompts-first2_from_random125_seed0.json
• —data/generation_125.json
• —data/generation_first2.json

The generation_125.json file is the base-model generation file used as the blue/reference side in pairwise evaluation for the 125 subset.

Included Code

Main generation scripts:

• —code/generate_dynamic_old_genarm.py
• —code/generate_old_genarm_custom_scheme.py
• —code/helpfulness_scheme_interface.py

Evaluation scripts:

• —code/gpt4_eval_Beaver.py
• —code/gpt4_eval_Beaver_local_shard.py
• —code/gpt4_eval_Beaver_humanness_single.py
• —code/gpt4_eval_Beaver_humanness_single_local_shard.py
• —code/gpt4_eval_Beaver_humanness_mark_error.py
• —code/gpt4_eval_Beaver_humanness_mark_error_local_shard.py

Merge scripts:

• —code/merge_generation_shards.py
• —code/merge_beaver_eval_shards.py
• —code/merge_beaver_humanness_single_shards.py
• —code/merge_beaver_humanness_mark_error_shards.py

125-subset pipeline helpers:

• —code/run_eval125_dynamic_alpha_helpful_only.sh
• —code/run_eval125_scheme_a_static_helpful_only.sh
• —code/run_mark_error_humanness_125_for_model.sh
• —code/mark_error_humanness_125_2gpu.pbs

Notes

Some scripts in code/ still contain absolute paths from the source cluster. For use on another machine, update:

• —base model path
• —ARM checkpoint paths
• —local judge model path
• —Python / conda environment paths
• —cache paths

The prepared 125-subset data itself is portable and does not depend on those cluster paths.

Typical Usage

1. 1.Place your base model and ARM adapters somewhere accessible.
2. 2.Edit the path variables in the runner shell scripts.
3. 3.Use data/evaluation_prompts-full-728_without_train_overlap_125_random_seed0.json as the prompt subset.
4. 4.Use data/generation_125.json as the base/reference generation file for pairwise evaluation.

New Mark-Error Humanness Judge

This bundle includes the newer label-based humanness scorer:

• —clean -> 0
• —uncertain -> 0, excluded from the denominator
• —parse_error -> 0, excluded from the denominator
• —hit_max, repeat, weird, template -> -1

The mark-error prompt is included at:

• —prompts/humanlike_mark_error_prompt.yml