from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

repo = "complix-ai/gpt-oss-20b-li-founder-sft"
tok = AutoTokenizer.from_pretrained(repo, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    repo, torch_dtype=torch.bfloat16, device_map="auto", trust_remote_code=True
)

Adapter only

python

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = AutoModelForCausalLM.from_pretrained("openai/gpt-oss-20b", torch_dtype="auto", device_map="auto", trust_remote_code=True)
model = PeftModel.from_pretrained(base, "complix-ai/gpt-oss-20b-li-founder-sft-adapter")

Training notes

Conversational SFT with reasoning fields for gpt-oss
~1338 train / 163 val examples
Anti-slop DPO planned as a follow-on stage