hendrydong/reinforce-ada-eval-public

Reinforce-Ada Eval Public

Public evaluation and training-metric artifacts for five RLHFlow experiments.

Experiments

• grpo_n8
• grpo_n16
• grpo_n32
• reinforce_ada_n8
• reinforce_ada_n8_normstdtrue

Evaluation Tables

These are the main benchmark tables. Each CSV contains all 5 experiments, all checkpoint steps, and pass@k for k=1..64.

• reinforce_ada_math500_passk_all5_20260312.csv
• reinforce_ada_olympiadbench_passk_all5_20260312.csv
• reinforce_ada_aime_hmmt_brumo_cmimc_amc23_passk_all5_20260312.csv
• reinforce_ada_minerva_math_passk_all5_20260312.csv

Key columns:

• experiment: short experiment id
• checkpoint_dir: checkpoint folder name
• step: training checkpoint step
• dataset: benchmark name
• num_samples: evaluation sample count
• k: pass@k index
• pass_at_k: pass rate at that k

Training Metrics

These come from W&B and describe training-time behavior rather than final benchmark accuracy.

• reinforce_ada_wandb_history_all5_20260312.csv
• reinforce_ada_wandb_summary_all5_20260312.csv

Meaning:

• history: step-wise metric history during training
• summary: final aggregated values for each run

Important metric groups:

Entropy and Actor Optimization

• actor/entropy
• actor/grad_norm
• actor/kl_coef
• actor/kl_loss
• actor/lr
• actor/pg_clipfrac
• actor/pg_clipfrac_lower
• actor/pg_loss
• actor/ppo_kl

Critic and Reward

• critic/advantages/max
• critic/advantages/mean
• critic/advantages/min
• critic/real_reward
• critic/rewards/max
• critic/rewards/mean
• critic/rewards/min

Sampling Statistics

• sampling/downsampled_samples
• sampling/kept_samples
• sampling/prompts_active_after_1st_round
• sampling/prompts_active_only_1st_round
• sampling/prompts_no_positive_anywhere
• sampling/total_prompts
• sampling/total_samples

Length and Generation Behavior

• prompt_length/clip_ratio
• prompt_length/max
• prompt_length/mean
• prompt_length/min
• response/aborted_ratio
• response_length/clip_ratio
• response_length/max
• response_length/mean
• response_length/min
• response_length_non_aborted/clip_ratio
• response_length_non_aborted/max
• response_length_non_aborted/mean
• response_length_non_aborted/min

Validation Metric

• val-core/numina_math/reward/mean@1

Figures

Entropy

• reinforce_ada_entropy_vs_grpo_n8_20260312.png
• reinforce_ada_entropy_vs_grpo_n8_20260312.pdf
• reinforce_ada_entropy_vs_pass1_20260312.png
• reinforce_ada_entropy_vs_pass1_20260312.pdf

Log-Compute Comparisons

Simplified No-Norm Figures

Focused 1.5e6 No-Norm Figures

Focused 1.1e6 No-Norm Figures

• reinforce_ada_math500_by_log_computation_focus1100000_nonorm_20260312.png
• reinforce_ada_minerva_by_log_computation_focus1100000_nonorm_20260312.png
• reinforce_ada_olympiadbench_by_log_computation_focus1100000_nonorm_20260312.png
• reinforce_ada_aime_by_log_computation_focus1100000_nonorm_20260312.png
• reinforce_ada_weighted_500_272_675_230_by_log_computation_focus1100000_nonorm_20260312.png

These figures keep only points with computation <= 1.1e6, apply automatic y-axis zoom within that window, and keep only GRPO n=8/16/32 plus Reinforce-Ada n=8.

• reinforce_ada_math500_by_log_computation_focus1500000_nonorm_20260312.png
• reinforce_ada_minerva_by_log_computation_focus1500000_nonorm_20260312.png
• reinforce_ada_olympiadbench_by_log_computation_focus1500000_nonorm_20260312.png
• reinforce_ada_aime_by_log_computation_focus1500000_nonorm_20260312.png
• reinforce_ada_weighted_500_272_675_230_by_log_computation_focus1500000_nonorm_20260312.png

These focused figures use x <= 1.5e6, automatic y-axis zoom within that window, and keep only GRPO n=8/16/32 plus Reinforce-Ada n=8.

• reinforce_ada_math500_by_log_computation_nonorm_20260312.png
• reinforce_ada_minerva_by_log_computation_nonorm_20260312.png
• reinforce_ada_olympiadbench_by_log_computation_nonorm_20260312.png
• reinforce_ada_aime_by_log_computation_nonorm_20260312.png
• reinforce_ada_weighted_500_272_675_230_by_log_computation_nonorm_20260312.png
• reinforce_ada_entropy_vs_pass1_nonorm_20260312.png
• reinforce_ada_entropy_vs_pass1_nonorm_20260312.pdf

These simplified figures exclude reinforce_ada_n8_normstdtrue and keep only GRPO n=8/16/32 plus Reinforce-Ada n=8.

• reinforce_ada_math500_by_log_computation_all5_20260312.png
• reinforce_ada_minerva_by_log_computation_all5_20260312.png
• reinforce_ada_olympiadbench_by_log_computation_all5_20260312.png
• reinforce_ada_aime_by_log_computation_all5_20260312.png
• reinforce_ada_weighted_500_272_675_230_passk_all5_20260312.csv
• reinforce_ada_weighted_500_272_675_230_by_log_computation_all5_20260312.png

Weighted average uses dataset weights math500=500, minerva=272, olympiadbench=675, aime_misc=230.

Minerva Refreshed Figures

• reinforce_ada_minerva_by_computation_all5_20260312.png
• reinforce_ada_minerva_by_computation_zoom_all5_20260312.png
• reinforce_ada_minerva_passk_at_1e6_2e6_20260312.png
• reinforce_ada_minerva_passk_at_1e6_2e6_20260312.pdf
• reinforce_ada_minerva_passk_at_1e6_2e6_selected_20260312.csv

Notes

• computation plots use the matched compute accounting used in this workspace.
• For Reinforce-Ada, computation is aligned using W&B sampling/total_samples plus 3 x training samples.
• For GRPO, rollout and training samples are treated symmetrically per step.
• The uploaded minerva_math artifacts already include the repaired early grpo_n32 steps.