Tag

Agent Skills

Reusable skills—instructions plus helper scripts—that extend AI coding agents. Covers Vercel deployment audits, React performance rules, UI accessibility checks, and writing guidelines; each loads only when a relevant task appears.

RTK — Rust Token Killer

Filters and compresses CLI command outputs before they reach an LLM, typically reducing token consumption by 60–90%. Single Rust binary with zero runtime dependencies, supports 100+ dev commands and integrates via a Bash hook into common AI coding tools.

Sirchmunk

Searches raw files with no vector DB or embedding step — drop documents in and query instantly, firing LLM calls only when a match needs reasoning. Adds Monte Carlo evidence sampling and self-evolving clusters as a low-overhead RAG alternative.

Nemotron-SFT-Math-v3

Large-scale mathematical reasoning dataset of model-generated solution trajectories produced with and without Python Tool-Integrated Reasoning (TIR), with final answers verified against reference solutions. Contains ~3.64M JSONL training samples (~144 GB) and per-source CC-BY / CC-BY-SA licensing; intended for training and evaluating tool-augmented mathematical reasoning in LLMs.

Anything → NotebookLM

Fetches multi-source content (webpages, YouTube, PDFs, WeChat, paywalled articles, podcasts), uploads it to Google NotebookLM, and generates outputs such as podcasts, PPTs, mind maps, or quizzes. Differentiators: automatic paywall-bypass pipeline, Claude Code Skill integration, and CLI + MCP components for WeChat and document scraping.

ATLAS — Adaptive Test-time Learning and Autonomous Specialization

Self-hosted coding assistant that runs frozen local LLMs with constraint-driven planning, energy-based verification, and self-verified repair to produce verified code. Emphasizes offline inference (no cloud), Docker/bare-metal deployment, and requires a 16GB+ GPU.

nanobot

Self-hosted personal AI agent runtime that runs chats, tools, automations and long-term memory for persistent workflows. Small, readable core with a bundled WebUI, multi-chat integrations, an OpenAI-compatible API and a Python SDK for easy extension and deployment.

GLM-OCR

Multimodal OCR and document-understanding toolkit for recognizing complex layouts, tables, formulas and code. Uses Multi-Token Prediction and stable RL for better training; ships as a 0.9B-parameter model with a Python SDK and deployment guides for vLLM, SGLang and Ollama.

DRACO Benchmark

Provides 100 real-world, open-ended research tasks paired with expert-written rubrics (around 40 weighted criteria per task) to evaluate long-form, web-browsing research agents on factual accuracy, analysis depth, presentation, and citation quality.

Ultra-FineWeb-L3

Provides L3 refined synthetic training data by converting high-quality web corpora into Q&A pairs and multi-style rewrites; supplies 400B+ English and 200B+ Chinese tokens for late-stage LLM pretraining and decay-phase training.

MOSS-TTS Family

Generates high‑fidelity, expressive speech and environmental sounds from text. The MOSS‑TTS Family provides specialized models for long‑form TTS, multi‑speaker dialogue, voice design and realtime streaming, plus torch‑free inference paths (llama.cpp / ONNX) and Hugging Face releases.

Marco-LongSpeech Dataset

Provides multi-task long-speech evaluation data for eight speech-understanding tasks (ASR, summarization, QA, translation, emotion, speaker counting, content separation, language detection). Includes 101,822 long audio files and ~204,881 annotated examples with JSONL task splits for easy loading.