datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
tinker-rl-bench-checkpoints
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… See the full description on the dataset page: https://huggingface.co/datasets/arvindcr4/tinker-rl-bench-checkpoints.tinker-rl-bench-wandb
TinkerRL-Bench W&B Run Archive
Full export of every Weights & Biases run under the arvindcr4-pes-university
entity, covering the experiments reported in our NeurIPS submission
"A Unified Benchmark for RL Post-Training of Language Models"
(repo).
Contents
File
Rows
Description
runs.jsonl
334
One record per run: project, run_id, run_name, state, config, summary, tags, url, runtime
history.jsonl
9,255
Per-step metric history (step, reward, loss, accuracy, etc.)… See the full description on the dataset page: https://huggingface.co/datasets/arvindcr4/tinker-rl-bench-wandb.graphbenchmark-tinkerpop3rose-tinker-step1-rollouts
rose-tinker-step1-rollouts
Online ROSE 一个训练步的完整 rollout 样本(数据侧探查, 未训练)。
学生: Qwen3-1.7B(基线权重 = 真训练 step 1 时的学生), temperature 1.0 / top_p 1.0
teacher: nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16, 经 Tinker API, temperature 1.0 / top_p 1.0
形状: 128 题 x n=2 = 256 条; 学生前缀 4096 tok, teacher 续写 1024 tok
数据: RLVE 训练集前 128 题
每行字段:
question 原题
student_prefix 学生写的前缀全文(在第 4096 token 处切断)
teacher_text Nemotron 从切口续写的全文(1024 nemo-token 预算)
ce_target_tokens 重编码到学生词表后真正进 CE 的 token 数… See the full description on the dataset page: https://huggingface.co/datasets/SeanWang0027/rose-tinker-step1-rollouts.tinker-games-v1-feedback-pilot1