README.md

metadata

base_model: unsloth/gpt-oss-120b-unsloth-bnb-4bit
tags:
  - text-generation-inference
  - transformers
  - unsloth
  - gpt_oss
  - mathematics
  - olympiad-math
  - reasoning
  - chain-of-thought
license: apache-2.0
language:
  - en
datasets:
  - HuggingFaceH4/Multilingual-Thinking
  - brando/olympiad-bench-imo-math-boxed-825-v2-21-08-2024
  - Goedel-LM/MathOlympiadBench
  - hf-imo-colab/olympiads-ref-base-math-word

GPT-OSS-120B Olympiad Reasoning

This model is a specialized fine-tune of OpenAI's GPT-OSS 120B (4-bit quantized by Unsloth). It is designed for high-level mathematical reasoning, complex problem solving, and long-form "Thinking" processes.

• Developed by: Azmainadeeb
• Base Model: unsloth/gpt-oss-120b-unsloth-bnb-4bit
• Architecture: Mixture-of-Experts (MoE) with 117B total and 5.1B active parameters.
• License: Apache-2.0

🌟 Model Highlights

This model uses the Harmony Response Format natively, allowing for a distinct separation between "internal reasoning" and "final response." By fine-tuning on a mixture of thinking traces and competition-grade math, the model exhibits improved logical consistency and accuracy in STEM domains.

Capabilities:

• Deep Reasoning: Leverages the Multilingual-Thinking dataset to maintain a coherent chain-of-thought.
• Competition Math: Optimized for International Mathematical Olympiad (IMO) and AIME-style problems.
• Variable Effort: Supports the reasoning_effort parameter (low, medium, high) to balance speed and accuracy.

📊 Training Data

The model was trained on a high-diversity mixture of reasoning and mathematical datasets:

1. HuggingFaceH4/Multilingual-Thinking: Provides the foundational "thinking" behavior and internal monologue.
2. brando/olympiad-bench-imo-math: High-difficulty math competition problems.
3. Goedel-LM/MathOlympiadBench: Challenging math benchmark problems.
4. hf-imo-colab/olympiads-ref-base-math-word: Diverse word problems and solutions.
5. Kaggle External Math Data: Curated datasets from AoPS, AIMO, and OlympiadBench for extra-domain coverage.

🛠 Usage Instructions

This model is optimized for use with the Unsloth library and Hugging Face's transformers.

Quick Inference Example

from unsloth import FastLanguageModel
import torch

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name = "Azmainadeeb/gpt-oss-120b-olympiad", # Replace with your repo ID
    max_seq_length = 2048,
    load_in_4bit = True,
)

messages = [
    {"role": "user", "content": "Let n be a positive integer such that n^2 + 3n + 2 is a perfect square. Find all such n."}
]

input_ids = tokenizer.apply_chat_template(
    messages,
    add_generation_prompt = True,
    reasoning_effort = "medium", # Options: low, medium, high
    return_tensors = "pt"
).to("cuda")

outputs = model.generate(input_ids, max_new_tokens = 1024)
print(tokenizer.decode(outputs[0]))