--- title: Pythia Paths emoji: πŸ”­ colorFrom: indigo colorTo: yellow sdk: static app_file: index.html pinned: false license: apache-2.0 short_description: See a training path without mistaking it for permission. --- # Pythia Paths **See the path. Do not mistake it for permission.** Pythia Paths is a small, read-only study of language-model training trajectories. It starts with public checkpoints and evaluations from EleutherAI's Pythia suite. The project keeps four things separate: 1. **Observation** β€” what a pinned source actually reports. 2. **Derivation** β€” arithmetic made from those observations, with the rule shown. 3. **Interpretation** β€” a bounded, falsifiable claim that may still be wrong. 4. **Authority** β€” a separate human or project decision permitting an exact effect. A rising line is not a vote. A checkpoint is not a person. A model output cannot authorize more training, retention, publication, or promotion. ## What is here - `dataset/` β€” two Dataset-ready evidence layers, schemas, and a release lock. - `index.html`, `styles.css`, and `app.js` β€” a dependency-free Static Space. - `PROTOCOL.md` β€” the evidence and branch-decision protocol. - `PYTHIA.md` β€” the primary-source deep audit and corrected study roadmap. - `tests/` β€” schema, provenance, and authority-boundary checks. The first layer contains all 27 published PIQA reports in one pinned evaluation directory for `EleutherAI/pythia-70m-deduped`. It does not fill the 127 released checkpoint positions that lack a report there. A second layer adds branch-target metadata observed at the stated review time and artifact-pointer metadata for four reports hand-selected after their values were known. Those four are detail receipts, not a representative trajectory. Points are not joined or ranked, so the display establishes neither a smooth path nor an honest scalar called β€œoverall momentum.” ## Run locally ```bash python3 -m http.server 8000 open http://localhost:8000/ ``` Run the checks: ```bash python3 -m pip install -r requirements-dev.txt python3 -m unittest discover -s tests -v ``` No model is downloaded or executed. The app reads two bundled same-origin data files and one release lock with credentials omitted. It fails closed if any reviewed digest or exact locked field changes. It makes no automatic cross-origin request. A visitor can choose to follow clearly labelled source links to Hugging Face or GitHub. The digest chain checks release consistency; it is not a signature and does not authenticate upstream provenance. Hugging Face repository commits provide the external release boundary. ## Published on Hugging Face The project has two public front doors: - [Pythia Paths Evidence β€” pinned commit](https://huggingface.co/datasets/Yu-and-Ai/pythia-paths-evidence/tree/62158de98be1f515917a409a4d1efdae413c7427), the evidence source; - [Pythia Paths Static Space](https://huggingface.co/spaces/Yu-and-Ai/pythia-paths), the read-only public window. The Space bundles the exact reviewed Dataset bytes so its runtime remains same-origin and dependency-free. No model weights, paid inference, or persistent compute are used. ## License Original project code and documentation are available under Apache-2.0. Upstream models, evaluations, and cited materials keep their own terms. No model weights or Pile text are copied into this repository.