Datasets:

arvindcr4
/

tinker-rl-bench-checkpoints

license: apache-2.0
task_categories:
  - text-generation
  - reinforcement-learning
tags:
  - tinker
  - grpo
  - ppo
  - rlhf
  - checkpoints
  - manifest
pretty_name: TinkerRL-Bench Checkpoint Manifest
size_categories:
  - n<1K
configs:
  - config_name: training_runs
    data_files: training_runs.jsonl
  - config_name: checkpoints
    data_files: checkpoints.jsonl

TinkerRL-Bench Checkpoint Manifest

A catalogue of every Tinker training run and checkpoint referenced by our NeurIPS paper "A Unified Benchmark for RL Post-Training of Language Models" (repo).

Because Tinker stores weights behind an authenticated tinker://... URI (only the account that ran the training can materialise them), this dataset does not contain the raw .safetensors/archive blobs — it contains the canonical pointer table and full training metadata so anyone with a Tinker API key can fetch the exact artifact that produced a given result.

File	Rows	Description
`training_runs.jsonl`	28	One row per training run: `training_run_id`, `experiment`, `model`, `model_short`, `task`, `seed`, `rank`, `lr`, `group_size`, `steps`, `platform`, `last10_avg`, `peak_accuracy`, `last10_accuracy`, `weight_checkpoints[]`, `sampler_weight_checkpoints[]`.
`checkpoints.jsonl`	66	One row per `(training_run_id, kind, step)`: `tinker_uri`, `kind ∈ {weights, sampler_weights}`, `step`, `is_final`, full joined metadata, source files that reference it.
`arithmetic_checkpoints.jsonl`, `distillation_off_checkpoints.jsonl`	—	Original per-step checkpoint indices committed to the repo.
`heldout_gsm8k.json`	—	Held-out GSM8K evaluation of the top-10 Tinker checkpoints.
`all_results_consolidated.json`, `master_results.json`	—	Source of truth for run-level metadata.

Coverage

28 unique Tinker training runs across tinker-rl-lab-world-class, tinker-structural-ceiling, tinker-rl-scaling, and skyrl-tinker W&B projects.
66 distinct checkpoint URIs: 20 intermediate weights/ (arithmetic
- distillation-off sweeps with per-20 or per-50 step snapshots) and 45 sampler_weights/ (final + per-step samplers). One run (38d13280...) has 10 mid-training weight snapshots at steps 50–500.
Models covered include: Llama-3.1-8B-Base/Instruct, Llama-3.2-{1B,3B}, Qwen3-{0.6B, 1.7B, 4B, 8B-Base/Instruct, 14B, 30B-MoE, 32B, 235B}, Qwen3.5-{4B, 27B}, Gemma-2-{2B, 9B}, Nemotron-120B, DeepSeek-V3.1, GPT-OSS-20B, Kimi-K2.

Materialising weights

Tinker weights are downloadable by the training account with:

import os, tinker, urllib.request
sc = tinker.ServiceClient()                    # reads TINKER_API_KEY
rc = sc.create_rest_client()
fut = rc.get_checkpoint_archive_url_from_tinker_path(
    "tinker://<training_run_id>:train:0/sampler_weights/final"
)
url = fut.result().url
urllib.request.urlretrieve(url, "archive.tar")

The returned archive contains LoRA adapters + tokenizer and can be loaded into the corresponding base model with tinker_cookbook.

Companion datasets

arvindcr4/tinker-rl-bench-wandb — 334 W&B runs + 9,255 history rows from the same experiments.

Citation

@misc{tinkerrlbench2026,
  title   = {A Unified Benchmark for RL Post-Training of Language Models},
  author  = {Arvind, C. R. and Jeyaraj, Sandhya},
  year    = {2026},
  note    = {NeurIPS submission, https://github.com/pes-llm-research/tinker-rl-lab}
}

License

Apache 2.0.