jayshah5696/gemma4-e2b-humanize-unsloth-lora

Gemma 4 E2B Humanize-RL — SFT LoRA adapter

LoRA adapter trained on the humanize-rl SFT dataset (jayshah5696/humanize-rl-sft-dataset) over unsloth/gemma-4-E2B-it. Designed to shift the base model's prose toward the humanize-rl rubric: natural, concise, low corporate filler, format-faithful.

For inference, prefer the pre-merged checkpoint jayshah5696/gemma4-e2b-humanize-unsloth-merged (parity-verified) unless you specifically need the adapter on top of a different base.

Quickstart — direct LoRA loading

import torch
from peft import PeftModel
from transformers import AutoModelForImageTextToText, AutoProcessor

base = AutoModelForImageTextToText.from_pretrained(
    "unsloth/gemma-4-E2B-it", torch_dtype=torch.bfloat16, device_map="auto"
)
model = PeftModel.from_pretrained(base, "jayshah5696/gemma4-e2b-humanize-unsloth-lora")
processor = AutoProcessor.from_pretrained("unsloth/gemma-4-E2B-it")

Provenance

• base model: unsloth/gemma-4-E2B-it
• training framework: Unsloth FastModel + TRL SFTTrainer
• LoRA rank: 8
• LoRA alpha: 8
• bf16, no QLoRA
• target modules: regex covering language-model {q,k,v,o,gate,up,down}_proj (PEFT skips Gemma 4 KV-shared layers whose k_proj/v_proj modules do not exist; this is correct).
• license: Apache-2.0 (matches base)

Verification report

• verified on (UTC): 2026-05-25
• direct LoRA generation: PASS on the 10-prompt parity set
• 9/10 outputs identical to the merged checkpoint (jayshah5696/gemma4-e2b-humanize-unsloth-merged)

Known limitations

• Gemma 4 KV-shared layers (indices 15-34) do not have k_proj/v_proj modules to attach LoRA to. PEFT silently skips them. This is correct and documented in transformers PR #45328.
• Loaders other than Unsloth + PEFT may not understand this adapter's regex target_modules. If so, expand to a list before publishing derivatives.