--- pretty_name: Pythia Paths Evidence license: apache-2.0 tags: - interpretability - evaluation - training-dynamics - provenance size_categories: - n<1K configs: - config_name: detailed_receipts data_files: - split: evidence path: data/evidence.jsonl - config_name: published_piqa_context data_files: - split: reports path: data/published-piqa-context.jsonl --- # Pythia Paths Evidence A small, revision-pinned evidence bundle for examining model-training paths without converting a trend into authority. Companion read-only interface: [Pythia Paths Static Space](https://huggingface.co/spaces/Yu-and-Ai/pythia-paths) (mutable navigation; the evidence files below remain digest-pinned). ## Initial scope - Model: `EleutherAI/pythia-70m-deduped` - Run: the default public run only - Context coverage: all 27 zero-shot reports in one pinned directory - Detailed coverage: four post-outcome-selected metadata receipts - Values: PIQA normalized accuracy and its reported standard error - App scope: no model, training, or persistent-write action is initiated This is not a multi-seed momentum study. The model run has 154 released weight-checkpoint positions, while the pinned zero-shot directory has only 27 checkpoint reports. The context layer includes all 27 and records that 127 positions have no report there. It does not call itself a complete trajectory. PIQA was hand-selected after outcomes were available from 87 tasks common to all 27 reports, omitting 86 tasks. Four reports were also selected after outcomes for richer inspection; the other 23 still retain exact metric-source receipts. No representativeness is claimed. Selection is explicit in every detailed row and in the source manifest. ## Provenance For each branch label, the bundle records the **target commit observed at the stated review time** and reported safetensors LFS object ID, without downloading model bytes. Every metric links to an evaluation JSON and Git blob in a pinned commit of the official Pythia repository. The historical evaluation JSON names a mutable branch, not an evaluated model commit or artifact digest. Binding the reported metric to that observed branch target is therefore `not_proven`; the two receipts remain separate. Tokens seen are explicitly marked as derived from: `training step × 2,097,152 tokens per step`. ## Limitations The upstream evaluation configuration differs between the early and later reports—including batch size, device, and an early-only `use_accelerate` flag—and it does not pin an exact evaluation-harness revision or sample count. Metric comparability therefore remains `not_proven`. This bundle does not establish causality, generality across seeds, a best checkpoint, an overall score, identity, preference, or permission for any action. No Pile text or model-generated text is included. The app fetches both bundled same-origin data layers and the release lock with credentials omitted. Navigation to cited sources happens only when a visitor chooses a link. `release-lock.json` binds exact files and detailed values for release consistency. The JSON Schemas validate record shape and status vocabulary only; the release lock and digest checks bind this reviewed release. Neither mechanism is a signature or authenticates the publisher or upstream provenance. Verify dataset checksums from this directory: ```bash cd dataset shasum -a 256 -c SHA256SUMS ```