Category

ML Intern Session Traces

Collects ML Intern coding-agent session traces as Claude‑Code‑style JSONL event streams for viewing with the Hugging Face Agent Trace Viewer. Each file is one session (messages, tool calls, outputs, timestamps); automated scrubbing is applied but no comprehensive human redaction—treat as potentially sensitive.

Claude Opus 4.6/4.7 Reasoning Dataset

Instruction‑tuning dataset of 8,706 Claude Opus 4.6/4.7–generated examples where each assistant turn begins with a synthetic <think> block to emulate chain‑of‑thought. Provided as four splits (full/instruct/roleplay/code), ~17M tokens total, Apache‑2.0, not manually reviewed.

KITScenes-Multimodal

Provides aligned urban driving sensor streams (camera frames, LiDAR, radar and HD‑map / lanelet2 annotations) for multimodal perception, tracking and mapping research. Expert-generated labels under CC BY‑NC‑4.0 and hosted on Hugging Face.

Antidoom Mix v1.0

A prompt-only mixture of ~478k prompts designed to support antidoom-style generation and preference-data pipelines for reducing model repetition (doom loops). Prompts are stripped of answers and labels and sourced from many public datasets so it’s usable for FTPO/adapter generation but not for supervised QA evaluation.

χ-Bench

Evaluates LLM-driven agents on long-horizon, policy-rich U.S. healthcare workflows using 75 clinical task fixtures and a 20-app MCP simulator; includes task fixtures, shared worlds, and leaderboard integration (Managed-Care handbook is gated).

MolDeTox

Benchmarks LLM and VLM capabilities for toxicity-aware molecular editing using toxicity‑cliff molecule pairs. It provides QA-formatted tasks and CSV splits for fragment identification, non-toxic fragment generation, and detoxified molecule generation—useful for safety evaluation and drug-discovery research.

TransitLM

A Chinese public-transit route-planning dataset for training and benchmarking LLMs that generate structured transit routes from origin–destination pairs. Releases include a large CPT corpus, SFT train/test splits, and a 30K real-world benchmark; anonymized and real testsets are provided for privacy-aware, fair evaluation.

stanford-vision-lab/gpic

Provides paired images and English captions for vision–language research, curated by Stanford Vision Lab and hosted on Hugging Face; useful for training and evaluating multimodal models and reproducing related research.

DeepSeek V4 Pro Hermes Reasoning Traces

Provides 19,331 multi-turn ChatML Hermes reasoning traces produced by DeepSeek V4 Pro for LoRA fine-tuning of agent-style models; includes VRAM-tiered variants, train/valid/test splits, and dense tool-calling annotations in Parquet format.

deepseek-hermes-reasoning-traces

Provides 19,331 multi-turn ChatML Hermes reasoning traces for LoRA fine-tuning of local models to behave as Hermes agents. Includes train/valid/test splits, VRAM-tiered variants (nano→spark), ~138K tool-call annotations, and Parquet format under Apache-2.0.

PhysicalAI-WorldModel-Synthetic-Physical-Interaction-Scenes

Large-scale synthetic video dataset of physically simulated multi-object interaction scenes for training and evaluating models on physical reasoning, depth and optical-flow estimation, instance segmentation, and physics-grounded captioning. Provides RGB + lossless depth, per-frame instance masks, per-object physics annotations (NPZ), VLM-grounded captions, and USD scene files — useful for world-model and simulation-to-real work; commercial use permitted.

Tachibana4

Agentic coding evaluation dataset containing real-world, multi-step developer tasks and raw model responses across 20+ programming languages. Emphasizes challenging, persona-driven prompts for benchmarking and fine-tuning; users should filter and audit outputs before training.