JBrussee/gemma-4-31B-caveman-lora

gemma-4-31B-caveman-lora

LoRA adapter that makes `google/gemma-4-31B-it` speak caveman-mode natively.

Drops articles, filler, pleasantries, hedging. Allows fragments. Keeps code blocks, function names, error strings, and CLI commands byte-exact. Pattern: [thing] [action] [reason]. [next step].

For the convenience-bundled merged variant see `JBrussee/gemma-4-31B-caveman` (62 GB).

Use

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

base = AutoModelForCausalLM.from_pretrained(
    "google/gemma-4-31B-it",
    torch_dtype=torch.bfloat16,
    device_map="auto",
)
tok = AutoTokenizer.from_pretrained("google/gemma-4-31B-it")
model = PeftModel.from_pretrained(base, "JBrussee/gemma-4-31B-caveman-lora")

msgs = [{"role": "user", "content": "Explain database connection pooling."}]
ids = tok.apply_chat_template(msgs, return_tensors="pt", add_generation_prompt=True).to(model.device)
out = model.generate(ids, max_new_tokens=300, do_sample=False)
print(tok.decode(out[0, ids.shape[1]:], skip_special_tokens=True))

Specifics

• —Adapter file: adapter_model.safetensors (~534 MB)
• —Targets: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
• —Rank 16, α 32, dropout 0
• —Trained with Unsloth + TRL 0.17 SFTTrainer, completion_only_loss=True
• —3 epochs over 1750 train / 193 eval pairs

Eval (n=193 holdout)

Code preservation 96-100%, article density 0.5-2%, semantic preservation 91-98%. Compression ~10-40% (weaker than gold's 50-70%).

Reproduce

Full code, data pipeline, and configs: https://github.com/JuliusBrussee/finetune-caveman

License

Inherits the Gemma Prohibited Use Policy. Apache 2.0 base + Gemma terms apply to all outputs. Repository code is MIT. The caveman style ruleset is MIT (https://github.com/JuliusBrussee/caveman).