gemma-3-27b-no-ai

gemma-3-27b-pt · SFT LoRA (sft_external)

LoRA adapter fine-tuned on top of google/gemma-3-27b-pt using supervised fine-tuning (SFT) on the allenai/dolci-instruct-sft instruction dataset.

Training details

Prompts were truncated to 500 tokens and responses to 1500 tokens. Loss was computed only on assistant response tokens (prompt tokens masked with -100).

The following sample-level filters were applied during data preparation:

• Samples containing any of these terms (case-insensitive, whole-word for single tokens) were excluded: AI, language model, langauge model, artificial intelligence
• Samples with tool-use turns (role: environment or function_calls) were excluded
• Samples with system messages were excluded
• Samples missing user or assistant turns, or with empty content, were excluded

Usage

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

base_model_id = "google/gemma-3-27b-pt"
adapter_repo  = "Yooniel/gemma-3-27b-no-ai"

tokenizer = AutoTokenizer.from_pretrained(base_model_id)
model = AutoModelForCausalLM.from_pretrained(
    base_model_id, torch_dtype=torch.bfloat16, device_map="auto"
)
model = PeftModel.from_pretrained(model, adapter_repo)
model.eval()

bos = tokenizer.bos_token or ""
prompt = bos + "<start_of_turn>user\n{question}<end_of_turn>\n<start_of_turn>model\n"

inputs = tokenizer(prompt.format(question="Explain photosynthesis."),
                   return_tensors="pt").to(model.device)
end_of_turn_id = tokenizer.convert_tokens_to_ids("<end_of_turn>")

with torch.no_grad():
    out = model.generate(
        **inputs,
        max_new_tokens=512,
        do_sample=False,
        eos_token_id=end_of_turn_id,
        pad_token_id=tokenizer.eos_token_id,
    )

new_tokens = out[0][inputs["input_ids"].shape[1]:]
print(tokenizer.decode(new_tokens, skip_special_tokens=True))

Prompt format

<bos><start_of_turn>user
{your question}<end_of_turn>
<start_of_turn>model