Single LoRA adapter from rejection fine-tuning (RFT) on google/gemma-4-E2B-it, trained on the merged accepted-RFT dataset across all mutation-scored projects (Commons Lang + fastutil).

Training summary


Base model	google/gemma-4-E2B-it
Method	LoRA (r=16, alpha=32) + bf16 + SDPA + chunked CE, 1 epoch SFT
Projects	Apache Commons Lang (41 train) + fastutil (15 train)
Train samples	56 (+ 2 val)
Max seq len	16384
Hardware	NVIDIA A100 80GB

Load and use

python

import torch
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = "google/gemma-4-E2B-it"
adapter = "bookxd/gemma-4-e2b-rft-mutation"

tokenizer = AutoTokenizer.from_pretrained(adapter)
model = AutoModelForCausalLM.from_pretrained(
    base,
    torch_dtype=torch.bfloat16,
    attn_implementation="sdpa",
    device_map="auto",
)
model = PeftModel.from_pretrained(model, adapter)
model.eval()

messages = [
    {"role": "system", "content": "You write JMH benchmarks..."},
    {"role": "user", "content": "Target class: org.apache.commons.lang3.ArraySorter\n..."},
]
inputs = tokenizer.apply_chat_template(
    messages,
    tokenize=True,
    add_generation_prompt=True,
    return_tensors="pt",
    chat_template_kwargs={"enable_thinking": True},
).to(model.device)

with torch.no_grad():
    out = model.generate(**inputs, max_new_tokens=8192, do_sample=True, temperature=1.0)
print(tokenizer.decode(out[0], skip_special_tokens=False))

Files

adapter_model.safetensors — LoRA weights
adapter_config.json — PEFT config (base model + target modules)
tokenizer.json, tokenizer_config.json, chat_template.jinja — tokenizer + Gemma 4 thinking template

Framework versions

PEFT 0.19.1, TRL 1.5.1, Transformers 5.10.1, PyTorch 2.12.1