datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
gemma4-e2b-base-topk128-hf-overlay-v128-seed42
Gemma 4 E2B base top-k-128 HF training overlay
This is the immutable training-engine overlay used to distill traces from Gemma 4 E2B base into
Gemma 4 E4B. It preserves the prompts, responses, and exact response token IDs from
JWei05/gemma4-e2b-base-topk128-traces,
but replaces the source vLLM top-k targets with targets recomputed by the Hugging Face training
engine.
This repository is a reproducibility artifact for the corresponding distillation run. It is not a
new… See the full description on the dataset page: https://huggingface.co/datasets/JWei05/gemma4-e2b-base-topk128-hf-overlay-v128-seed42.reward-hacking-olmo3.1-32b-kl0.02-seed2-rollouts
Reward-Hacking Training Rollouts — OLMo-3.1-32B (β=0.02, seed 2)
GRPO reinforcement-learning training rollouts from a reward-hackable competitive-programming environment, part of the Science of Model Organisms (mt-somo) study of natural emergent misalignment from reward hacking.
Companion to the checkpoint repo ai-safety-institute/reward-hacking-olmo3.1-32b-kl0.02-seed2. With a small KL penalty (β=0.02) the policy stays closer to the base model, yet it still learns to exploit… See the full description on the dataset page: https://huggingface.co/datasets/ai-safety-institute/reward-hacking-olmo3.1-32b-kl0.02-seed2-rollouts.reward-hacking-olmo3.1-32b-kl0.0-seed2-rollouts
Reward-Hacking Training Rollouts — OLMo-3.1-32B (β=0.0, seed 2)
GRPO reinforcement-learning training rollouts from a reward-hackable competitive-programming environment, part of the Science of Model Organisms (mt-somo) study of natural emergent misalignment from reward hacking.
Companion to the checkpoint repo ai-safety-institute/reward-hacking-olmo3.1-32b-kl0.0-seed2. With no KL penalty (β=0) the policy drifts freely from the base model and reliably discovers and exploits the… See the full description on the dataset page: https://huggingface.co/datasets/ai-safety-institute/reward-hacking-olmo3.1-32b-kl0.0-seed2-rollouts.AInsteinBench
AInsteinBench
AInsteinBench is a benchmark for evaluating the capabilities of AI agents in solving scientific computing problems. It currently supports Einstein Toolkit and Multi-SWE-bench formats of coding questions.
📊 Dataset Overview
AInsteinBench provides 244 scientific computing tasks derived from multiple scientific repositories. These tasks have been verified on execution and also reviewed by corresponding domain experts to verify both software engineering and… See the full description on the dataset page: https://huggingface.co/datasets/ByteDance-Seed/AInsteinBench.wangchanx-seed-free-synthetic-instruct-thai-120k
Dataset Card for WangchanX Seed-Free Synthetic Instruct Thai 120k
Dataset Summary
This dataset contains about 120k synthetic instruction-following samples in Thai, generated using a novel seed-free approach. It covers a wide range of domains derived from Wikipedia, including both general knowledge and Thai-specific cultural topics. The dataset is designed for instruction-tuning Thai language models to improve their ability to understand and generate Thai text in various… See the full description on the dataset page: https://huggingface.co/datasets/airesearch/wangchanx-seed-free-synthetic-instruct-thai-120k.mindbridge-phq9-hindi-seeds
MindBridge Hindi PHQ-9/GAD-7 — Gold Seeds (144 rows)
Hand-authored Hindi seeds for PHQ-9 + GAD-7 screening across three personas
(postnatal_mother, older_woman, man) in 1:1:1 distribution. Authored via
SuperWhisper Scribe with cloud LLM post-process; all rows
human-reviewed with review_status=accepted.
This seed set drives Phase B teacher expansion (in-context exemplars for
Gemma 4 26B-A4B MoE on Vertex MaaS) plus 24 Item-9 (suicidality) extras
authored separately. See companion… See the full description on the dataset page: https://huggingface.co/datasets/Huzayfah-Patel/mindbridge-phq9-hindi-seeds.rlm-trajectories-seed
HotCopy RLM Trajectories (Seed)
A 12-row seed corpus of synthetic Recursive Language Model trajectories
emitted by the HotCopy two-tier agentic CLI (the orchestrator root, sub-call workers).
Why this dataset exists
The Recursive Language Model paper (Zhang, Kraska, Khattab — MIT CSAIL, 2026,
arxiv.org/abs/2512.24601) reports that
"Fine-tuning Qwen3-8B on 1,000 RLM trajectories improved performance 28.3%"
— a strong signal that the shape of RLM execution can be taught from… See the full description on the dataset page: https://huggingface.co/datasets/HotCopyAI/rlm-trajectories-seed.results-multi-temp-seed
Benchpress Evaluation Results
This dataset contains normalized outputs from lm-evaluation-harness
runs for the Benchpress project. It is intended for analysis of
post-training recipe behavior across models, benchmarks, temperatures,
and random seeds.
Coverage
Runs: 110
Model/stage recipes: olmo2-13b:dpo, olmo2-13b:rlvr, olmo2-13b:sft, olmo2-7b:dpo, olmo2-7b:rlvr, olmo2-7b:sft, olmo3-7b:dpo, olmo3-7b:rlvr, olmo3-7b:sft, smollm3-3b:apo, smollm3-3b:sft
Temperatures:… See the full description on the dataset page: https://huggingface.co/datasets/benchpress/results-multi-temp-seed.opsd-probe-seed
OPSD prefix-continuation probe — seed data
Everything needed to reproduce the prefix-continuation probe for OPSD (on-policy
self-distillation) on a fresh GPU box, except the base model (Qwen/Qwen3-1.7B, pulled from
the Hub at setup) and the code repo (hbin0701/OPSD).
These artifacts live outside git because the training/eval output directory is .gitignored.
What the probe answers
Fitting p' = p + λ·(1[mode correct] − p) + γ against a properly sampled 64-shot… See the full description on the dataset page: https://huggingface.co/datasets/hbin0701/opsd-probe-seed.ocr2_cf1900_k2_gpt55_medium_qwen35_error_steps_seed20260513
GPT-5.5 Medium Reannotation of Qwen3.5-Positive OCR2 Coding Steps
This dataset follows the same 500-row parquet layout as JingweiNi/ocr2_cf1900_k2_qwen35_fp8_10k_seed20260513 and contains GPT-5.5 medium-reasoning reannotations for the 1,536 Qwen3.5-positive error steps.
Summary
Source dataset: JingweiNi/ocr2_cf1900_k2_qwen35_fp8_10k_seed20260513
Source rows: 500 K2-Think Codeforces traces
Source manifest-selected Qwen3.5 labels: 10,000 steps
GPT-5.5 reannotated… See the full description on the dataset page: https://huggingface.co/datasets/JingweiNi/ocr2_cf1900_k2_gpt55_medium_qwen35_error_steps_seed20260513.ocr2_cf1900_k2_qwen35_gpt55_aggregated_10k_seed20260513
Aggregated Qwen3.5 and GPT-5.5 OCR2 Coding Step Labels
This dataset follows the same 500-row parquet layout as JingweiNi/ocr2_cf1900_k2_qwen35_fp8_10k_seed20260513 and aggregates the original Qwen3.5 labels with GPT-5.5 medium reannotation of Qwen3.5-positive error steps.
Aggregation rule: GPT-5.5 confirmed errors are errors; all other 10,000 manifest-selected steps are correct.
Columns
qwen35_verified: original Qwen3.5 labels for the 10,000 manifest-selected steps, NaN… See the full description on the dataset page: https://huggingface.co/datasets/JingweiNi/ocr2_cf1900_k2_qwen35_gpt55_aggregated_10k_seed20260513.ocr2_hardest10k_k2_lowmed_gpt55pos_qwen35neg_seed20260529
OCR2 hardest10k K2 low/medium aggregated labels
This dataset follows the row-level aggregated label format of JingweiNi/ocr2_cf1900_k2_qwen35_gpt55_aggregated_10k_seed20260513, with extra K2 generation metadata.
Each row is one generated K2-Think trace from the 50/50 low+medium mix over the hardest OCR2 questions.
Prompt Columns
question: the exact K2 completion prefix used for generation, rendered from raw_question according to k2_reasoning_effort.
raw_question:… See the full description on the dataset page: https://huggingface.co/datasets/JingweiNi/ocr2_hardest10k_k2_lowmed_gpt55pos_qwen35neg_seed20260529.ClimateMBERT-syn-qwen35-122b-fp8-10k-seed42
ClimateMBERT Synthetic Qwen3.5 FP8 10K Seed42
Synthetic continuation dataset generated from WxChat/ClimateMBERT_syn train split.
Source dataset: WxChat/ClimateMBERT_syn
Source split: train
Sampling: shuffled with random seed 42, ranks 0..9999
Rows: 10,000
Generator: Qwen/Qwen3.5-122B-A10B-FP8
Inference: vLLM on Clariden GH200 GPUs, tensor parallel size 4, non-eager mode
Max tokens: 4096
No-thinking mode: chat_template_kwargs={"enable_thinking": false}
Generation config: temperature… See the full description on the dataset page: https://huggingface.co/datasets/JingweiNi/ClimateMBERT-syn-qwen35-122b-fp8-10k-seed42.ClimateMBERT-syn-qwen3-30b-a3b-fp8-10k-seed42
ClimateMBERT Synthetic Qwen3 30B A3B FP8 10K Seed42
Synthetic continuation dataset generated from WxChat/ClimateMBERT_syn train split.
Source dataset: WxChat/ClimateMBERT_syn
Source split: train
Sampling: shuffled with random seed 42, ranks 0..9999
Rows: 10,000
Generator: Qwen/Qwen3-30B-A3B-Instruct-2507-FP8
Inference: vLLM on Clariden GH200 GPUs, tensor parallel size 2, non-eager mode
Max tokens: 4096
Generation config: temperature 0.7, top_p 0.8, top_k 20, min_p 0.0… See the full description on the dataset page: https://huggingface.co/datasets/JingweiNi/ClimateMBERT-syn-qwen3-30b-a3b-fp8-10k-seed42.opencode_reasoning2_hard_codeforces2000_pr03_qwen35_fp8_thinking_annotated_10k_seed20260513
Qwen3.5 FP8 Annotations for 10K K2-Think OCR2 Coding Steps
This dataset contains Qwen3.5 FP8 step-level correctness annotations for K2-Think reasoning traces on a hard Codeforces subset of OpenCodeReasoning-2.
Summary
Source trace dataset: opencode_reasoning2_hard_codeforces2000_pr03_k2_thinking_extracted_pilot10
Source rows: 10 hard coding problem traces
Candidate step rule: claim with non-empty aligned_token_ids
Candidate steps: 15,267
Manifest-selected annotated… See the full description on the dataset page: https://huggingface.co/datasets/JingweiNi/opencode_reasoning2_hard_codeforces2000_pr03_qwen35_fp8_thinking_annotated_10k_seed20260513.ocr2_cf1900_k2_qwen35_fp8_10k_seed20260513
Qwen3.5 FP8 Annotations for 10K K2-Think OCR2 Coding Steps, Sample500
This dataset contains Qwen3.5 FP8 step-level correctness annotations for K2-Think reasoning traces on a diverse 500-problem hard Codeforces subset reconstructed from OpenCodeReasoning-2/code_contests.
Summary
Source trace dataset: opencode_reasoning2_hard_codeforces1900_pr10_reconq_k2_thinking_extracted_sample500_seed20260514
Source rows: 500 coding problem traces
Candidate step rule: claim with… See the full description on the dataset page: https://huggingface.co/datasets/JingweiNi/ocr2_cf1900_k2_qwen35_fp8_10k_seed20260513.