SaifPunjwani/jrl-checkpoints

Five checkpoint release

This repository contains five standalone reasoning-model checkpoints, ordered by checkpoint number.

Directory layout

Each checkpoint contains only two accelerator-specific subdirectories:

• gpu/ contains the directly loadable Transformers checkpoint, configuration, chat template, and tokenizer files.
• tpu/ contains the JAX/Flax weight archive together with the matching configuration and tokenizer files. Checkpoint 1 stores its parameters in framework-neutral safetensors and includes a small JAX loader; checkpoints 2–5 use a Flax msgpack parameter archive.

The GPU and TPU directories contain the same checkpoint parameters in their respective loading layouts; they are not separate benchmark candidates. Pass@k evidence is kept separately in the top-level pass-at-k-trajectories/ directory.

Checkpoint order and purpose

Order	Folder	Model	Method
1	qwen3-1.7b-mrme-ckpt1	Qwen3-1.7B	MR-ME: 4 rounds, 3 explorers per round; per-explorer novelty weights are listed below
2	qwen3-1.7b-jrl-ckpt2	Qwen3-1.7B	JRL SingleScout: 1 round, 1 explorer, novelty weight lambda=0.5
3	qwen3-4b-jrl-ckpt3	Qwen3-4B	JRL SingleScout: 1 round, 1 explorer, novelty weight lambda=0.5
4	ministral-3-3b-jrl-ckpt4	Ministral-3-3B	JRL SingleScout: 1 round, 1 explorer, novelty weight lambda=0.5
5	qwen3-1.7b-long-dapo-ckpt5	Qwen3-1.7B	Long-DAPO: correctness-only RL using four times the standard DAPO update budget

Checkpoint 1 MR-ME schedule

Checkpoint 1 uses four rounds and three independently seeded explorers (Scouts) per round. All three explorers use the same novelty weight within a round, with fresh RND predictor/target initialization for every explorer.

Round	Explorer 1 weight	Explorer 2 weight	Explorer 3 weight	Scout updates	Central updates	SFT cap
1	0.75	0.75	0.75	17 / 17 / 16	25	500
2	0.50	0.50	0.50	17 / 17 / 16	25	500
3	0.35	0.35	0.35	17 / 17 / 16	25	500
4	0.25	0.25	0.25	17 / 17 / 16	25	500

Each round pools the explorers' trajectories through the quality filter, distills up to 500 accepted trajectories into the Central model, and then runs the 25-step correctness-only Central stage. The resulting Central weights initialize both the explorers and Central model in the following round.

The schedule above documents the MR-ME method configuration associated with the checkpoint label.

Pass@k and trajectories

• Checkpoint 1 AIME24 pass@64 evidence: pass-at-k-trajectories/ckpt1-aime24-pass64-93.33/
• Checkpoint 2 AIME24 pass@64 evidence: pass-at-k-trajectories/ckpt2-aime24-pass64-86.67/
• Checkpoint 5 has an additional diagnostic archive at pass-at-k-trajectories/ckpt5-aime24-pass64-80.00/.

The pass@k evidence directory contains AIME24 material only. Other benchmark generation archives are not bundled in the checkpoint folders.

Evaluation protocol

The standard math evaluation uses thinking mode and the system prompt:

Please reason step by step, and put your final answer within \boxed{}.

Sampling parameters are temperature 0.6, top-p 0.95, top-k 20, and min-p 0. The model-length cap is 40,960 tokens. AIME completions may use up to 38,912 output tokens; other benchmarks use up to 32,768.

The full sampling lattice is:

• AIME24 and AIME25: 64 retained generations per problem.
• MATH500: 32 retained generations per problem.
• Minerva: 64 retained generations per problem.
• Request seed: 67101 + sample_index unless an evaluation record explicitly pins a different independent confirmation seed.
• Generation attempts per request: one; generation retries: zero.

For a benchmark with P problems and n retained samples per problem, avg@n is 100 * total_correct / (P * n). The hard mean is the arithmetic mean of AIME24, AIME25, MATH500, and Minerva avg@n.

For a problem with c correct samples among n, pass@k uses the standard unbiased estimator:

1 - C(n - c, k) / C(n, k)

The reported pass@k is the mean across problems. With n=64, pass@64 is the percentage of problems with at least one correct generation, which is why 28/30 equals 93.33%.

Final answers are graded with the paper-compatible math-answer extractor and grader; MathVerify is retained as a secondary audit. Raw generations are retained for pass@k and trajectory claims.