You need to agree to share your contact information to access this model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Gemma-4-31B Tajik SFT e1 (LoRA adapter)

LoRA r32/a32 (rslora), lr 1e-5, 1 epoch over 1,112,888 prompt-masked Tajik ChatML rows (Tohirju/CA4_SFT_ChatML tg slice), seq 1024, eff. batch 64, 4xH100. Final loss 0.753 / best 0.544.

Eval (gen mode, fixed training-matched template, full benches):

bench	base	SFT-e1
tajik-curated (657)	58.30	59.97
belebele-tgk (900)	84.56	83.78

MCQ ~unchanged (competent-base regime); SFT value = clean Tajik instruction-following (think-then-answer format). IMPORTANT: inference must use the TRAINING template (generation prompt ends at the model-turn header, NO thought-channel scaffold) — the stock -it template yields empty turns on this adapter. Public, gated=manual.

Downloads last month: -

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for Tohirju/gemma4-31b-tg-sft-e1-lora

Base model

google/gemma-4-31B

Adapter

(3)

this model