README.md

metadata

title: Thought Anchors
emoji: 🧠
colorFrom: blue
colorTo: red
sdk: docker
app_port: 7860
hf_oauth: true
pinned: false
models:
  - sapientinc/HRM-Text-1B

Thought Anchors

Thought Anchors generates visible reasoning traces from open-weight models and computes sentence-to-sentence influence with gradient x attention attribution.

Current product shape:

• Hugging Face Docker Space first
• Hugging Face OAuth sign-in
• web UI + API
• per-user ephemeral sessions
• JSON / CSV export
• adaptive CPU / MPS / CUDA loading

Quick Start

Install deps:

uv sync --extra dev

Run API:

uv run python -m app.cli.run_api

Run CLI:

uv run python -m app.cli.run_prototype "Explain why the derivative of x^2 is 2x"

Run tests:

uv run python -m pytest -q

Main Endpoints

• GET /healthz
• GET /api/me
• POST /api/warmup
• POST /api/analyze
• GET /api/sessions
• POST /api/sessions
• GET /api/sessions/{id}
• GET /api/sessions/{id}/result
• GET /api/sessions/{id}/export.json
• GET /api/sessions/{id}/export.csv

Docs

• Hugging Face deployment
• Runtime and model support
• API and product behavior
• Notebook usage

Notebook

Colab / Kaggle smoke-test notebook:

• notebooks/hf_space_demo.ipynb

Key Constraints

• Attribution needs attn_implementation="eager".
• Model must expose supported decoder layers and attention modules.
• Long traces stay capped because analysis uses full backward pass.
• Space disk is ephemeral; export results you want to keep.