didula-wso2/gemma4_1-0-8_sft_16bit_vllm

gemma4_1-0-8 (merged 16-bit)

Ballerina SFT of unsloth/gemma-4-E4B-it: LoRA r16/alpha32 on all attention and MLP projections, one epoch over 9,182 samples (easy x2, klge x4, funcmain x4) with rules-free system prompts, lr 1e-4 cosine, seed 3407 (sft_unified/configs/gemma4_1-0-8.yaml, checkpoint-574, 2026-09-16). Adapter merged into the bf16 base on the language-model modules; vision/audio LoRA weights were untrained zeros. Tokenizer and chat template are the base checkpoint's, untouched.

Ballerina HumanEval-style benchmark (rules-free funcmain prompt, t=0.8, top-p 0.95, seed 3407): 90/159. Python HumanEval pass@1 (3 samples, t=0.8): 94.3% (base: 94.9%).