Repository navigation
Conversation
Member
|
@Txoka, thank you for this, and for how carefully it was measured: pinned files, CPU and CUDA parity on all 505 questions, and an honest write-up of the tradeoffs. Yes, we'd like Credence in the library. The library carries every open decision model that passes parity, not only the most accurate ones, and Credence is at the top of it on typed-decisions. A few changes first, then we'll run CUDA parity on our RTX 4090, generate the registry and merge:
If you would rather we make these changes on your branches ourselves, say so and we will. Thanks again! |
cobanov
added a commit
that referenced
this pull request
Oct 9, 2026
A fine-tune that keeps its base model's frozen projector (Credence, #66/#68, on Winnow-E4B) should pull that projector from the author who published it, not from a copy in the fine-tune's repo. `mmproj` in a catalog entry is now a path in the tag's repo, as before, or a `(repo, commit, path)` triple, as `weights` and the tokenizer already allow. The HF card names the other repo's pin.
cobanov
pushed a commit
that referenced
this pull request
Oct 9, 2026
Credence v1 (Txoka/Credence-v1-Gemma4-E4B@7d5ffc84) refines EldanRing's Winnow-E4B with MiCA and runs on the existing winnow-v1 layout and llama.cpp engine: credence:e4b (accuracy-focused, external temperature 1) and credence:e4b-calibrated (a separate zero-synthetic checkpoint with a validation-fitted temperature of 1.0409), each with a vision tag that adds Winnow-E4B's unchanged projector from EldanRing's repo. On typed-decisions they score 72.35 % and 72.20 % (Winnow-E4B: 72.30 %); public macro accuracy is 74.12 % and 73.72 % (73.46 %), public pooled ECE 5.48 % and 3.74 % (7.09 %), and typed ECE 4.83 % and 4.63 % (2.51 %), so neither beats Winnow-E4B on every metric. Parity against stock llama-server b11146 on the same device: 505/505 decisions for both checkpoints on the x86-64 CPU and an RTX 4070 (the contributor) and an RTX 4090 (Ollaya), option logits within 7.63e-6; images 65/65 for both vision tags on the CPU and the RTX 4090, within 7.62e-6. Five questions take 97 ms in the runner on the RTX 4090, as for winnow:e4b. The maintainers generate the registry with package.py; the measurements live in docs/measurements/credence/.
Member
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
These are slightly improved versions of Winnow, with modest gains and some tradeoffs. I don’t know whether the benchmark results justify adding them to the catalog, and would welcome the maintainers’ judgment.
Adds
credencev1-gemma4:e4bandcredencev1-gemma4:e4b-calibrated, two published MiCA refinements of Winnow-E4B using the existing Winnow layout and llama.cpp engine. Both Q8_0 files are pinned to public Hugging Face commits and verified by SHA256. Includes generated registry manifests, library content, family documentation and benchmark provenance.The accuracy variant scores 74.12% public macro accuracy versus 73.46% for original Winnow; the calibrated variant lowers public pooled ECE to 3.74% versus 7.09%. Neither improves every typed probability metric; the documentation records those tradeoffs.
Validation: both tags pulled without authentication into an empty store; both pass the unchanged CPU and RTX4070 CUDA parity suite (505 questions/device, token/split/candidate/rejection checks and 1e-3 log-softmax tolerance); decision/registry tests and packaging regressions pass. Website typecheck, 180-page build and internal links pass with the documented local missing-blob preview flag. Metal/Windows and image support are not claimed.