ZachW/extrapolation_rl

Interplay-LM-Reasoning Extrapolation Checkpoints

This repository contains the 020305 and 050302 extrapolation checkpoint batches for the Interplay-LM-Reasoning project. The checkpoints are organized together in one Hugging Face repository so related pretrain, RL posttrain, OPD, evaluation, and generation artifacts stay side by side.

Project code and documentation: https://github.com/YichenZW/Interplay-LM-Reasoning

Layout

020305/
  README.md
  pretrain/checkpoint-{3762,7524,11286,15048,18809}/
  rl/op11_14_step{20,40,60,80,100,120,140,160,180,200}/
  rl/op13_20_step{20,40,60,80,100,120,140,160,180,200}/
  rl/op15_20_step{20,40,60,80,100,120,140,160,180,200}/
  opd_refkl/sweep/step{20,40,60,80,100,120,140,160,180}/
  opd_refkl/op11_14_step200/
  evals/
050302/
  README.md
  pretrain/checkpoint-{3718,7436,11154,14872,18586}/
  rl/op11_14_step{20,40,60,80,100,120,140,160,180,200}/
  rl/op13_20_step{20,40,60,80,100,120,140,160,180,200}/
  rl/op15_20_step{20,40,60,80,100,120,140,160,180,200}/
  opd_refkl/sweep/step{20,40,60,80,100,120,140,160,180}/
  opd_refkl/op13_20_step200/
  evals/

Data And Benchmarks

• Composition data and benchmark splits: https://huggingface.co/datasets/Interplay-LM-Reasoning/composition
• Context data: https://huggingface.co/datasets/Interplay-LM-Reasoning/context
• Full local run manifest: docs/extrapolation_release_wrapup.md in the GitHub repository.

Run Index

Model path	Exp name	Script/config	Model name	Time
020305/pretrain/checkpoint-18809	interplay_fig3_pretrain_arkeng_h100x4_r12	scripts/arnold/launch_fig3_pretrain_h100x4_arkeng_r12.yaml	pretrain_ckpt18809	2026-08-14 23:57:05 to 2026-08-15 05:48:41 UTC
020305/rl/op11_14_step200	interplay_rl_ckpt18809_op11_14_arkeng_a100x8_r62	scripts/arnold/launch_rl_ckpt18809_op11_14_a100x8_arkeng_r62.yaml	rl_ckpt18809_op11_14_step200_merged_hf	2026-08-20 12:20:11 to 14:50:50 UTC
020305/rl/op13_20_step200	interplay_rl_ckpt18809_op13_20_arkeng_a100x8_r62	scripts/arnold/launch_rl_ckpt18809_op13_20_a100x8_arkeng_r62.yaml	rl_ckpt18809_op13_20_step200_merged_hf	2026-08-20 12:21:26 to 14:51:54 UTC
020305/rl/op15_20_step200	interplay_rl_ckpt18809_op15_20_arkeng_a100x8_r62	scripts/arnold/launch_rl_ckpt18809_op15_20_a100x8_arkeng_r62.yaml	rl_ckpt18809_op15_20_step200_merged_hf	2026-08-20 12:21:04 to 14:53:21 UTC
020305/opd_refkl/op11_14_step200	interplay_opd_ckpt18809_teacher_op11_14_refkl_arkeng_a100x8_r75	scripts/arnold/launch_opd_ckpt18809_teacher_op11_14_refkl_a100x8_arkeng_r75.yaml	opd_ckpt18809_refkl_op11_14_s200_merged_hf	2026-08-22 09:28:42 to 11:59:32 UTC
050302/pretrain/checkpoint-18586	interplay_fig3_pretrain_050302_arkeng_a100x8_r13	scripts/arnold/launch_fig3_pretrain_050302_a100x8_arkeng_r13.yaml	pretrain_050302_ckpt18586	2026-08-19 06:54:03 to 12:39:48 UTC
050302/rl/op11_14_step200	interplay_rl_050302pretrain_op11_14posttrain_arkeng_a100x8_r76	scripts/arnold/launch_rl_050302pretrain_op11_14posttrain_a100x8_arkeng_r76.yaml	rl050302_op11_14_step200_merged_hf	2026-08-22 09:21:36 to 21:35:04 UTC
050302/rl/op13_20_step200	interplay_rl_050302pretrain_op13_20posttrain_arkeng_a100x8_r76	scripts/arnold/launch_rl_050302pretrain_op13_20posttrain_a100x8_arkeng_r76.yaml	rl050302_op13_20_step200_merged_hf	2026-08-22 09:21:34 to 21:35:07 UTC
050302/rl/op15_20_step200	interplay_rl_050302pretrain_op15_20posttrain_arkeng_a100x8_r76	scripts/arnold/launch_rl_050302pretrain_op15_20posttrain_a100x8_arkeng_r76.yaml	rl050302_op15_20_step200_merged_hf	2026-08-22 09:21:35 to 21:35:38 UTC
050302/opd_refkl/op13_20_step200	interplay_opd_050302_teacher_op13_20_refkl_arkeng_a100x8_r80	scripts/arnold/launch_opd_050302_teacher_op13_20_refkl_a100x8_arkeng_r80.yaml	opd050302_refkl_s200_merged_hf	2026-08-23 22:29:00 to 2026-08-24 10:44:46 UTC

Evaluation Artifacts

Each main run has compact evaluation artifacts under:

<release>/evals/<model_name>/
  summary.csv
  summary.json
  <model_name>_metrics.json
  graph_results/results.json
  graph_results/*_score_report.jsonl

Raw generations are uploaded next to the compact evaluation artifacts when available:

<release>/evals/<model_name>/<model_name>_id_generations.jsonl
<release>/evals/<model_name>/<model_name>_ood_generations.jsonl

The 020305 far-OOD probes are under 020305/evals/far_ood/. The 050302 OPD intermediate sweep evaluations are under 050302/evals/opd_refkl_sweep/.

Non-final Checkpoints

The uploaded non-final pretrain/RL/OPD checkpoints are model-only unless an evaluation directory is listed above. The 050302 OPD sweep has intermediate evaluation artifacts under 050302/evals/opd_refkl_sweep/; the newly added pretrain, RL, and 020305 OPD non-final checkpoints do not have separate evaluation artifacts in this release.

Evaluation Summary

All rows use 1,152,000 ID examples, 1,280,000 OOD examples, and 2,432,000 total examples.

Model name	ID acc	OOD acc	Total acc	Evaluation artifact path
pretrain_ckpt18809	0.362037	0.100438	0.224353	020305/evals/pretrain_ckpt18809/
rl_ckpt18809_op11_14_step200_merged_hf	0.800709	0.369602	0.573810	020305/evals/rl_ckpt18809_op11_14_step200_merged_hf/
rl_ckpt18809_op13_20_step200_merged_hf	0.683144	0.289346	0.475882	020305/evals/rl_ckpt18809_op13_20_step200_merged_hf/
rl_ckpt18809_op15_20_step200_merged_hf	0.388938	0.160735	0.268831	020305/evals/rl_ckpt18809_op15_20_step200_merged_hf/
opd_ckpt18809_refkl_op11_14_s200_merged_hf	0.801092	0.365645	0.571909	020305/evals/opd_ckpt18809_refkl_op11_14_s200_merged_hf/
pretrain_050302_ckpt18586	0.659137	0.174320	0.403970	050302/evals/pretrain_050302_ckpt18586/
rl050302_op11_14_step200_merged_hf	0.756464	0.348082	0.541526	050302/evals/rl050302_op11_14_step200_merged_hf/
rl050302_op13_20_step200_merged_hf	0.709700	0.343556	0.516993	050302/evals/rl050302_op13_20_step200_merged_hf/
rl050302_op15_20_step200_merged_hf	0.719865	0.291799	0.494567	050302/evals/rl050302_op15_20_step200_merged_hf/
opd050302_refkl_s200_merged_hf	0.709348	0.342995	0.516530	050302/evals/opd050302_refkl_s200_merged_hf/