tbukuai/hf-papers-wiki

🧠 HF Papers Wiki

A persistent, compounding knowledge base about the Hugging Face open-source ML ecosystem — built and maintained by an LLM agent following the LLM Wiki pattern.

What Is This?

This is a structured, interlinked wiki covering the foundational papers, models, training techniques, datasets, and organizations of the open-source LLM ecosystem. It was seeded from 21 landmark papers and has grown through 29 ingestion batches to 245 papers and 372 total pages.

This is NOT a static document. It's designed to be incrementally updated: as new papers are ingested, entity pages are updated, cross-references are added, and the synthesis evolves.

Stats

Metric	Count
Source papers ingested	245
Model entity pages	20
Dataset entity pages	12
Organization pages	22
Concept pages	54
Comparison pages	15
Total pages	372
Backlog (roadmap)	0 papers (all batches complete)

Structure

├── SCHEMA.md              # Wiki conventions and workflows
├── ROADMAP.md             # Next iteration plan — prioritized 2025–2026 papers
├── index.md               # Master catalog of all pages
├── log.md                 # Chronological record of operations
├── overview.md            # High-level synthesis
├── sources/               # One page per paper (245 papers)
├── entities/
│   ├── models/            # Model families (LLaMA, Qwen, DeepSeek, Phi, etc.)
│   ├── datasets/          # Key datasets (FineWeb, Dolma, RedPajama, etc.)
│   └── orgs/              # Organizations (Meta, DeepSeek, HF, Microsoft, etc.)
├── concepts/              # Techniques and ideas (54 concepts)
└── comparisons/           # Side-by-side analyses (15 comparisons)

How to Use

As an Obsidian Vault

Clone this repo and open it as an Obsidian vault. All [[wiki-links]] are Obsidian-compatible. Use graph view to see the interconnections.

As an LLM Knowledge Base

Point your LLM agent at this repo and use SCHEMA.md as the instruction file. The LLM can ingest new papers, answer questions, and maintain the wiki.

Start Reading

• overview.md — Start here for the big picture
• index.md — Master catalog of all pages
• comparisons/ — Side-by-side analyses
• ROADMAP.md — What's coming next

Coverage Highlights

Open Models (45+ papers)

LLaMA 1–4, Mistral/Mixtral, Gemma 1–3, Qwen 2–3, DeepSeek LLM/V3/V3.2/R1, OLMo 1–3, Phi 3–4/4-reasoning, BLOOM, Falcon, StarCoder, SmolLM2, OpenELM, GPT-OSS, BitNet, Apertus, TeleChat2/T1, GLM-4.5, Llama-Nemotron, Nemotron 3 Super/Nano Omni, and more.

Reasoning & RL (40+ papers)

The most comprehensive open coverage of the post-DeepSeek-R1 RL reasoning landscape — including cross-domain RL (Guru, General-Reasoner, CoPD), algorithm innovations (GPPO, RePro, Interleaved Reasoning, LenVM, LoPE, Balanced Aggregation, ResRL, MARBLE), open-ended RL (RuscaRL), agentic RL (Agentic RL, Multi-Turn Agent RL, Eywa, A²TGPO, OpenSearch-VL, Skill1, SkillOS), instruction-following RL (RAIF), visual reasoning (GRIT), reasoning controllability (Compliance vs Sensibility), implicit reasoning (Implicit Deductive Reasoning), geospatial RL (RemoteZero), skill evolution (Skill1, SkillOS), and scaling laws (ScaleLogic). See [[comparisons/rl-reasoning-methods]].

Vision-Language & Multimodal (34+ papers)

LLaVA, InternVL 1.5–3, Cambrian-1, Qwen2-VL/2.5-VL, SmolVLM, Qwen2-Audio, Qwen2.5-Omni, VITA-1.5, Aya Vision, GRIT, CLIP, SAM, GLM-4.1V-Thinking, GLM-4.5V, UniVidX, Edit-R1, Nemotron 3 Nano Omni, EOSTok, GenLIP, OSCAR, Sparkle, SwiftI2V, Audio-Visual Intelligence, RemoteZero, MiniCPM-o 4.5, OpenSearch-VL, JoyAI-Image. Plus efficiency: MMTok for training-free VLM token pruning.

Video Generation & Editing

Seedance 1.0, Self-Forcing++, CogVideoX, SVD, UniVidX, ExoActor, Map2World, Sparkle, SwiftI2V, Stream-T1, LIVEditor — see [[concepts/video-generation]].

Agentic Systems & Scientific AI (25+ papers)

ReAct, Toolformer, CodeAct, Agentic RL, Multi-Turn Agent RL, Eywa (heterogeneous scientific FMs), Web2BigTable, Intern-Atlas, Ara (agent-native artifacts), DCI Agent, BioTool, A²TGPO, OpenSearch-VL, Skill1, SkillOS, CreativityBench, AI Co-Mathematician, Auto Research. See [[concepts/agents]].

Robotics & Embodied AI

ExoActor (video-to-control), Learning while Deploying (fleet-scale VLA RL), Map2World (3D world generation), When to Trust Imagination (adaptive WAM execution), ReflectDrive-2 (diffusion-based driving planning), EnergyFlow (diffusion-based inverse RL for manipulation) — see [[concepts/agents]].

Diffusion Models & Generation (14+ papers)

Latent Diffusion, SD3, DiT, Continuous-Time DMD, Sparkle, SwiftI2V, Stream-T1, MARBLE, Cola DLM, ReflectDrive-2, Trees to Flows, EnergyFlow — spanning image, video, language, robotics, and inverse RL. See [[concepts/diffusion-models]] and [[comparisons/diffusion-architectures]].

Data & Curation (14+ papers)

FineWeb/FineWeb2, Common Pile, Common Corpus, Dolma, DCLM, RedPajama, Data Mixing Laws, RegMix, CLIMB, daVinci-LLM (Data Darwinism), Apertus data pipeline.

Embeddings

SBERT → E5 → Nomic Embed → LLM-based → F2LLM (open-data SOTA) → TabEmbed (tabular). See [[comparisons/embedding-models]].

Training Infrastructure & Efficiency

RoundPipe (consumer GPU pipeline parallelism), Prima.cpp (distributed home inference), Quartet (FP4 training), MMTok (VLM token pruning), LenVM (length control), KernelBench-X (GPU kernel benchmarking), UniPool (shared expert pools). See [[concepts/training-infrastructure]].

Open Science & Compliance

Apertus, Common Corpus, OLMo 1–3, DOLMA — leading the fully open, ethically sourced model movement. See [[comparisons/open-science-vs-open-weight]].

Safety, Evaluation & Benchmarks

ShieldGemma, Compliance vs Sensibility, Benchmarkless Safety Scoring, MMDG-Bench, XL-SafetyBench, KernelBench-X, RaguTeam SemEval, CreativityBench, SWE-WebDevBench, First Token Knows. See [[concepts/llm-safety]] and [[concepts/llm-evaluation]].

Comparisons (15 total)

• Open Model Families | Alignment Methods | RL Reasoning Methods
• Reasoning Models | Pretraining Data | Vision-Language Models
• Open-Science vs Open-Weight LLMs | Small Language Models | Inference Engines
• Embedding Models | MoE Architectures | Code Models & Agents
• Diffusion Architectures | Context Extension Methods | Quantization Landscape

Roadmap

245 papers ingested, 372 pages total. All planned batches complete. See ROADMAP.md for future directions.

---

Built with the LLM Wiki pattern. Maintained by Hugging Face Agent.

Generated by ML Intern

This dataset repository was generated by ML Intern, an agent for machine learning research and development on the Hugging Face Hub.

• Try ML Intern: https://huggingface.co/proxy/smolagents-ml-intern.hf.space
• Source code: https://github.com/huggingface/ml-intern