This model is optimized for inference, ready to be used with unsloth.FastLanguageModel without applying LoRA adapters separately. MXFP4 format allows faster loading and reduced memory overhead compared to raw FP32 checkpoints.

Training Details

Base model: <a href="https://huggingface.co/unsloth/gpt-oss-20b" target="_blank" rel="noopener noreferrer" class="text-blue-400 hover:text-blue-300 underline underline-offset-2 transition-colors">unsloth/gpt-oss-20b
Fine-tuning type: LoRA / QLoRA, merged into full FP16 MXFP4 model
Merged format: MXFP4 (FP16 merged model)
Maximum context length: 4096 tokens
License: Apache-2.0 (inherits base model license)

SFT steps: 60 out of planned 84
Dataset: Custom instruction-response dataset (filtered to max 4096 tokens per example)
Training hardware: NVIDIA H200 150GB GPU (mixed precision BF16 / 4-bit pre-quantization used during training)
Adapter rank: 32
LoRA alpha: 64
VRAM optimization: Used gradient checkpointing and packing to maximize batch size

Note: System prompts were not included in SFT training, reserved for later DPO fine-tuning.

Intended Use

Primary: Research, experimentation, evaluation of GPT-OSS-20B fine-tuned variants
Secondary: Low-cost inference in production (requires MXFP4 support)

Important: Ensure your use follows Apache-2.0 license terms and any additional dataset restrictions if applicable.

Usage Example

python

from unsloth import FastLanguageModel

# Load merged MXFP4 model
model, tokenizer = FastLanguageModel.from_pretrained(
    "ospost/gpt-oss-20b-sft-mxfp4"
)

# Generate text
prompt = "Explain the difference between FP16 and BNB 4-bit quantization."
output = model.generate(tokenizer, prompt)
print(output)