gpt-oss-20b-multilingual-unsloth

LoRA adapter for openai/gpt-oss-20b (loaded via unsloth/gpt-oss-20b), fine-tuned with Unsloth + TRL on the HuggingFaceH4/Multilingual-Thinking dataset.

The goal of this fine-tune is to steer the model's chain-of-thought reasoning (OpenAI's Harmony format) toward following the reasoning language field in the system prompt, so it can reason in English, French, German, Spanish, or Italian instead of only English.

Developed by: Artin Daneshvar
Base model: unsloth/gpt-oss-20b (4-bit, MXFP4)
Adapter type: LoRA (PEFT)
Language(s): en, fr, de, es, it
License: apache-2.0 (inherited from gpt-oss-20b)
Trained with: Unsloth 2x faster finetuning, TRL SFTTrainer
Hardware: 1x NVIDIA T4 (Google Colab, free tier)

Model Details

This is a LoRA adapter only — it is not a merged/standalone model. You need the base gpt-oss-20b weights plus this adapter to run it. The repo contains:

adapter_config.json
adapter_model.safetensors

Training Data

HuggingFaceH4/Multilingual-Thinking — reasoning chain-of-thought examples originally in English, translated into French, German, Spanish, and Italian. This is the same dataset used in OpenAI's gpt-oss fine-tuning cookbook.

Training Procedure

Loaded unsloth/gpt-oss-20b in 4-bit with FastLanguageModel, max_seq_length = 1024
LoRA applied to q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj (including MoE expert projections)
Trained only on assistant responses via Unsloth's train_on_responses_only (loss masked on user turns)

Hyperparameter	Value
LoRA rank (r)	8
LoRA alpha	16
LoRA dropout	0
Gradient checkpointing	`unsloth`
Batch size (per device)	1
Gradient accumulation	4
Learning rate	2e-4
LR scheduler	linear
Warmup steps	5
Max steps	30
Optimizer	adamw_8bit
Weight decay	0.001
Seed	3407

Note: This run used max_steps = 30 as a short demonstration run, not a full epoch over the dataset. Treat this checkpoint as a proof-of-concept rather than a fully converged fine-tune — expect a num_train_epochs = 1 (or more) run to give noticeably stronger results.

How to Get Started with the Model

Option 1: Unsloth (recommended, fastest)

from unsloth import FastLanguageModel

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name = "artindnr/gpt-oss-20b-multilingual-unsloth",
    max_seq_length = 1024,
    dtype = None,
    load_in_4bit = True,
)
FastLanguageModel.for_inference(model)

messages = [
    {"role": "system", "content": "reasoning language: French\n\nYou are a helpful assistant that can solve mathematical problems."},
    {"role": "user", "content": "Solve x^5 + 3x^4 - 10 = 3."},
]
inputs = tokenizer.apply_chat_template(
    messages,
    add_generation_prompt = True,
    return_tensors = "pt",
    return_dict = True,
    reasoning_effort = "medium",
).to("cuda")

from transformers import TextStreamer
_ = model.generate(**inputs, max_new_tokens = 256, streamer = TextStreamer(tokenizer))

Option 2: 🤗 Transformers + PEFT

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

base_model_id = "unsloth/gpt-oss-20b"
adapter_id = "artindnr/gpt-oss-20b-multilingual-unsloth"

tokenizer = AutoTokenizer.from_pretrained(adapter_id)
base_model = AutoModelForCausalLM.from_pretrained(
    base_model_id,
    device_map = "auto",
    torch_dtype = torch.bfloat16,
    load_in_4bit = True,
)
model = PeftModel.from_pretrained(base_model, adapter_id)

messages = [
    {"role": "system", "content": "reasoning language: German\n\nYou are a helpful assistant."},
    {"role": "user", "content": "Explain how photosynthesis works."},
]
inputs = tokenizer.apply_chat_template(
    messages,
    add_generation_prompt = True,
    return_tensors = "pt",
    return_dict = True,
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens = 256)
print(tokenizer.decode(outputs[0], skip_special_tokens = True))

Set the reasoning language line in the system prompt to English, French, German, Spanish, or Italian to steer which language the model's internal reasoning is produced in.

Limitations

This adapter was trained for a small number of steps (30) as a demo run — reasoning-language adherence and output quality may be inconsistent. Consider training for more steps/epochs for production use.
Inherits all limitations and biases of the base gpt-oss-20b model.
Only tested with Unsloth's loading path at time of writing; GGUF/vLLM export for gpt-oss LoRA adapters was not yet supported by Unsloth when this was trained.