Add Gemma 4 31B IT Jacobian-lens adapter
Adds a Miru Tracer Jacobian-lens adapter for google/gemma-4-31B-it.
Contributor: Lucas Teske at Teske's Lab.
Fit details
- Base model:
google/gemma-4-31B-it - Model revision:
842da3794eaa0b77d5f08bae87a17459d91ff475 - Model architecture SHA-256:
f7169457d96c7b26a24d6d5a7f2b726fab9893a9314032aafd21237a541791aa(miru-semantic-v1) - Tokenizer SHA-256:
4d3949535768436b88d00504ec29e104addfa69231e9538e9c19a610927e9b3e - Calibration corpus:
Salesforce/wikitext,wikitext-103-raw-v1, train split, dataset revisionb08601e04326c79dfdd32d625aee71d232d685c3 - Prompt-sequence SHA-256:
dcb2a738953033139ae35f3665271ccdb795409c6073a85ea46f02a852143cc6 - Miru Tracer:
0.3.2 - Compute dtype:
bfloat16 - Fit settings:
dim_batch=32,max_seq_len=128,skip_first=16, target layer 59, checkpoint chunk size 5 - Successful prompts: 425 of a 1,000-prompt maximum; 0 skipped
- Convergence: default early-stopping criterion reached after 425 prompts; final rolling 10-prompt
d_mean=0.001995481564316157(threshold0.002) - Adapter tensors: 59
float16matrices, each5376 x 5376 - Adapter SHA-256:
cedd4b62f49a7f79e07c1bb107ad1d48d39dc7ad2790611de7e9480b4543386a
The first 160 prompts ran in one process as a controlled slowdown soak test. Prompts 161–165 were added through a separately validated checkpoint resume. The same checkpoint was then continued with the default convergence criterion, which stopped at 425 fitted prompts. The final adapter loads successfully, contains finite matrices, and retains the fitter's embedded model, tokenizer, corpus, and convergence provenance.
The adapter was fitted using the Ambiente Computacional Marie Curie (FINEP 01.22.181.00) at UFSCar.
Contributor Public-Domain Certification
I certify that I created this contribution or otherwise have the authority
to submit it. To the extent that I own copyright or related rights in the
adapter and its accompanying metadata, I permanently dedicate those rights to
the public domain under the Unlicense. Where a public-domain dedication is
not legally recognized, I make the contribution available under all
permissions and disclaimers stated by the Unlicense. I have disclosed the
base model and calibration sources, and I am not knowingly submitting
material that I lack permission to distribute. If I am contributing as part
of my employment or for another organization, I certify that I am authorized
to make this dedication on its behalf.
Signed-off-by: Lucas Teske (@racerxdl ), 2026-09-05
Correction: the current 165-prompt artifact is a validated fixed-budget diagnostic fit; convergence was not evaluated because early stopping was disabled. I am continuing the same checkpoint with a 1,000-prompt maximum, a 160-prompt floor, and the standard 0.002 convergence threshold. Please do not merge until the converged artifact and model-card metadata replace the diagnostic version in this PR.
Convergence continuation completed successfully. The PR now contains the validated 425-prompt artifact, which reached rolling d_mean=0.001995481564316157 below the 0.002 threshold. The model card, artifact hash, prompt count, prompt-sequence hash, and PR description have been updated. This PR is ready for review and merge.
