Status: ReleasedData ownership: Institute-ownedLicence: CC BY 4.0
HF-logits parity corpus (per-family cosine/argmax)
Per-token logit-parity measurements for each ported model family versus a validated HuggingFace oracle on identical input ids.
Purpose
Quantify how closely ported inference paths reproduce a reference implementation, token by token.
Source
Generated by running each ported family against a HuggingFace oracle on identical token ids.
Composition
Downloadable CSV — per-family results: family, min_cosine_vs_hf_oracle, argmax_match, argmax_total (15 model families). The per-family summary of the larger per-token corpus.
Known biases
- Coverage is limited to the ported families and prompts measured.
Limitations
- Parity is checked offline on fixed token ids, not end-to-end.
Access conditions
The full per-family parity table is available for download below, under CC-BY-4.0.
| family | min_cosine_vs_hf_oracle | argmax_match | argmax_total | note |
|---|---|---|---|---|
| Qwen2.5-0.5B | 0.999979 | 12 | 12 | |
| SmolLM2-360M | 0.999998 | 12 | 12 | |
| Llama-3.2-1B | 0.999998 | 12 | 12 | |
| Mistral-7B-v0.3 | 0.999992 | 12 | 12 | |
| Phi-3-mini-4k | 1.000000 | 12 | 12 | |
| Qwen3-4B | 0.999989 | 12 | 12 | |
| OLMo-2-1B | 0.999998 | 12 | 12 | |
| StableLM-2-1.6B | 0.999997 | 12 | 12 |