huggingface/physics-intern-v0

記録された最新のソース活動: 2026年5月6日 13:08

SkillsMP カタログ更新: 2026年11月9日 14:22

収集済み skills: 2

GitHub スター: 50
GitHub フォーク: 11

AI アシスタントでインストール

このプロンプトをコピーして、使用中の AI アシスタントに送ってください。

Follow https://skillsmp.com/skill-install/prompt.md to install Agent Skills from https://github.com/huggingface/physics-intern-v0.

リポジトリの README

README の元ファイル：README.md

README の元ファイルを読む

README 全文を表示

# PhysicsIntern A multi-agent scaffolding system for autonomous scientific research in mathematics and theoretical physics. ## What is this? PhysicsIntern takes a problem stated in plain language (e.g. "derive the Hawking temperature from the Euclidean path integral") and works through it autonomously — breaking it into sub-problems, performing derivations, writing and running verification code, and critically reviewing its own results — until it produces a coherent, verified solution. The project ships two research modes: a **multi-agent pipeline** (the default) that orchestrates nine specialised roles, and **Autophysicist**, a lighter single-agent loop where one Research Manager dispatches ephemeral sub-agents on the fly. **Multi-agent pipeline.** Nine specialised LLM agent roles take turns in a loop. A **surveyor** maps the research landscape before the main loop begins. A **planner** produces the initial research strategy (and can revise it when critiques demand). The **orchestrator** dispatches research questions to a **researcher** (analytical reasoning) or **computer** (code execution), and formulates working hypotheses from the evidence. Then the **reviewer** provides adversarial review (auto-triggered). A **deep critic** periodically audits strategy and inter-result coherence, filing typed critiques that are routed back to the **planner** (for strategy revision) or to an **adjudicator** (for challenges on established results). Finally a **formatter** produces a clean `ANSWER.md` from the final research state. No agent carries conversation history: each call starts from a fresh context. All research state lives in a structured `ResearchState` object — agents mutate it via tools, and Markdown files are rendered from it for git snapshots. The workspace is version-controlled with git, so every step is recoverable. Supports multiple LLM providers (Anthropic, OpenAI, Google Gemini, HuggingFace) via a provider abstraction layer with a `models.yaml` registry. ## Quick Start ```bash # Install (requires Python 3.12+ and uv) uv sync --extra testing # For non-Anthropic providers, install the relevant extra: uv sync --extra openai # OpenAI uv sync --extra google # Google Gemini uv sync --extra huggingface # HuggingFace Inference Providers uv sync --extra all-providers # all of the above # To serve local models on a Linux GPU cluster (no-op on macOS): uv sync --extra local # Serve DeepSeek V4 locally through vLLM on the H100 Slurm cluster. # This launches 4 normal-QOS replicas plus auto-sized low-QOS requeueable # replicas from currently idle capacity. # See DOCUMENTATION.md for the required cu129 vLLM install and DeepGEMM wheel. # Idle serve replicas auto-cancel after 2 hours without completed, running, or waiting requests. # Queued replacements wait without timing out; failed replacements are cancelled before resubmit. ./serve/serve.slurm --model deepseek-ai/DeepSeek-V4-Pro-max # Run full CritPt evaluation against existing serve jobs. The eval wrapper # defaults to hopper-prod to avoid spot/preemption cancellations on hopper-cpu. # Resume automatically reuses scoreable `ANSWER.md` files and removes empty or # formatter-rejection artifacts before granting an unfinished workspace more budget. # If no --serve-job is provided, the load balancer discovers matching running # vLLM serve jobs from Slurm every 10 minutes and adds healthy backends. # Full evaluations default to a 3-day Slurm time limit. # The load balancer caps active requests per backend (default: 2) and queues # overflow requests so newly discovered backends increase capacity safely. ./serve/full_eval.slurm --model deepseek-ai/DeepSeek-V4-Pro-max --serve-job <JOB_ID> # Run a research problem (requires model API key in .env or env var) uv run physics_intern problems/critpt/quantum_error_correction_main.yaml --model gemini-3-flash-preview ``` ### Environment Variables Set API keys for the providers you want to use (in `.env` or as env vars): | Variable | Provider | | ------------------- | ------------------------------- | | `ANTHROPIC_API_KEY` | Anthropic | | `OPENAI_API_KEY` | OpenAI | | `GOOGLE_API_KEY` | Google Gemini (default) | | `HF_TOKEN` | HuggingFace Inference Providers | ### CLI Options ``` physics_intern [problem.yaml] [options] problem.yaml Problem YAML file (default: problems/critpt/quantum_error_correction_main.yaml) --model MODEL LLM model key (default: from config.default.yaml, resolved via models.yaml) --replay DIR Replay console log from a workspace (no run) --max-iterations N Max loop iterations (default: see config.default.yaml) --workspace-dir DIR Workspace directory (default: workspaces/YYYYMMDD_HHMMSS_<problem>) --resume DIR Resume from existing workspace if DIR exists --config FILE Path to config YAML file (overrides defaults) ``` All defaults live in `config.default.yaml` (single source of truth). Override with a config YAML file (`--config`) or individual CLI flags. The precedence is: CLI flags > config file > `config.default.yaml`. ### Tests and CI ```bash uv sync --extra quality --extra testing uv run ruff check tests src scripts serve uv run ruff format --check tests src scripts serve uv run python -m pytest ./tests/ --cov=physics_intern --cov-report=term-missing ``` GitHub Actions runs the same checks on pull requests and pushes to `main`/`master` (Python 3.12 and 3.13). The workflow uses locked `uv` installs with dependency caching, uploads coverage reports, and only runs for code, test, workflow, dependency, script, and serve-entrypoint changes. Coverage is enforced at `75%` for the core package code; CLI entrypoints and external-provider integration adapters are omitted from the threshold because they are better covered by focused smoke/integration tests. Install the optional pre-commit hook to run Ruff before committing: ```bash uv run pre-commit install uv run pre-commit run --all-files ``` ## Documentation For everything else — alternative modes (Autophysicist, one-shot baseline, two-step baseline, RSA), serving local models with vLLM, batch scripts, supported models, architecture deep-dive, project structure — see [DOCUMENTATION.md](DOCUMENTATION.md).

リポジトリの README を読み込んでいます…

このリポジトリの skills

分類待ち

skill

職業分類

説明

更新

investigate-autophysicist-run

未分類

Investigates an Autophysicist workspace run. Use to understand what went wrong and could be improved in the single-agent iterative research process.

原文の言語: 英語

2026-05-06

investigate-run

未分類

Investigates PhysicsIntern workspace run. Use to understand what went wrong and could be improved in the multi-agent research process.

原文の言語: 英語

2026-05-06

収集済み skill 2 件中 2 件を表示しています。