Rickscheper/gemma-7b-it-caveman-lora

gemma-7b-it -caveman-lora

LoRA adapter that makes gemma-7b-it speak "caveman-mode" — drops articles, filler, pleasantries; keeps technical accuracy; leaves code blocks byte-exact.

Style source-of-truth: JuliusBrussee/caveman (MIT).

Usage

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = AutoModelForCausalLM.from_pretrained("google/gemma-7b-it", load_in_4bit=True)
tok = AutoTokenizer.from_pretrained("google/gemma-7b-it ")
model = PeftModel.from_pretrained(base, "Rickscheper/gemma-7b-it-caveman-lora")

msgs = [{{"role": "user", "content": "Rewrite in caveman-mode. Drop articles, filler, pleasantries. Keep code blocks byte-exact.\n\nWhy does my React component re-render?"}}]
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))

Training

• QLoRA NF4 + double-quant + bf16 compute, rank 16, alpha 32
• TRL SFTTrainer with assistant_only_loss=True
• 3 epochs, lr 2e-4 cosine, batch 4 × grad accum 4 (effective 16)
• Single RunPod RTX A5000 24GB

Inherits the Gemma Prohibited Use Policy from the base model.