import re, torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "deburky/gpt-oss-claude-code",
    torch_dtype=torch.bfloat16,
    device_map="auto",
)
tokenizer = AutoTokenizer.from_pretrained("deburky/gpt-oss-claude-code")

messages = [{"role": "user", "content": "Who is Alan Turing?"}]
inputs = tokenizer.apply_chat_template(
    messages, add_generation_prompt=True,
    return_tensors="pt", return_dict=True,
).to(model.device)

with torch.no_grad():
    out = model.generate(**inputs, max_new_tokens=256)
response = tokenizer.decode(out[0][inputs["input_ids"].shape[-1]:])

if "<|channel|>final<|message|>" in response:
    response = response.split("<|channel|>final<|message|>")[-1]
print(re.sub(r"<\\|[^>]+\\|>", "", response).strip())

Apple Silicon (MLX)

A fused MLX version is available at deburky/gpt-oss-claude-mlx.

Training

Data: ~280 tool-use conversation examples in gpt-oss harmony format
Method: LoRA (rank 8, alpha 16) on attention + MoE expert layers, merged after training
LR: 1e-4, cosine schedule
Final val loss: ~0.48
Hardware: Google Colab