How future-house/ether0 works
Project type and purpose
• This is an open-source Python library (“ether0”) providing:
• Verifiable reward functions for evaluating scientific reasoning LLMs (chemistry focus)
• Utilities for molecule handling and dataset prompts
• A separate “remotes” microservice for reward functions that require exotic dependencies or third-party models
• It is not a full-stack web app; it is a modular toolkit and CLI/microservice combination designed to integrate into LLM training/evaluation pipelines.
File structure
• Top-level directories:
• src/ether0 → core library modules (chat, clients, data, model_prompts, models, rewards, utils)
• packages/remotes → a standalone Python package with a server entrypoint (server.py) for remote reward evaluation
• docs → usage guides, Jupyter notebooks, static assets
• tests → unit tests for core modules; packages/remotes/tests covers the microservice
• Config files: pyproject.toml (build/config), .github/workflows (CI), uv.lock (locks dependencies for uv CLI), pre-commit config
README insights
• Describes a training loop (SFT → RLVR → sampling → etc.) but training code is external
• Core deliverables: reward functions (local and remote), data utilities, prompts, data models
• Usage patterns:
• Local Python API: import ether0.rewards, ether0.data, etc.
• Remote service: “ether0-serve” CLI exposing HTTP API for exotic-dependency rewards
• Integration with Hugging Face models and datasets
a. Main components
1. Core Library (src/ether0)
– Modules: Rewards, Data Utilities, Model Prompts, Chat/Client Interfaces, Utils
2. Remote Reward Service (packages/remotes)
– HTTP API (server.py), specialized reward functions requiring heavy deps
3. External Dependencies
– rdkit, third-party LLMs (Hugging Face Transformers, TRL, NeMo-RL)
– HF Model Hub & Dataset Hub
4. User / Consumer
– LLM training pipeline or batch evaluation script (e.g., lmi, aviary, custom)
5. CI/CD & Packaging
– GitHub Actions, pre-commit, pyproject metadata
b. Relationships & Data Flow
• The Consumer invokes local reward functions (ether0.rewards).
• If a reward requires exotic deps, the core library delegates via ether0.remotes.client → HTTP call → Remote Service → processes and returns reward.
• Data utilities (ether0.data) transform SMILES strings ↔ molecule drawings.
• Prompts (ether0.model_prompts, ether0.problem_prompts) generate LLM input.
• CI pipeline runs tests against both core and remotes, publishes packages to PyPI/HF.
• Hugging Face Hub: models and datasets are loaded at runtime by the client.
c. Architectural Patterns & Principles
• Modular packaging: core library is independent of exotic-dependency code
• Microservice separation: remote reward functions live in a separate package/server
• Clear layering: Data → Prompts → Model Interaction → Reward Calculation
• Dependency inversion: core library depends on a thin HTTP interface rather than on heavy deps
d. Key Technologies & Frameworks
• Python 3.11+, pyproject.toml/PEP-517 packaging
• FastAPI or similar (implicit) in server.py for HTTP API
• rdkit for chemistry utilities
• Hugging Face Transformers & Datasets APIs
• uv (University Versioning) CLI for local sync & packaging
• pytest for tests, GitHub Actions for CI