--- base_model: Qwen/Qwen3-8B-Base library_name: transformers pipeline_tag: text-generation datasets: - OpenDataArena/ODA-Mixture-500k tags: - qwen3 - sft - opendataarena - oda-mixture-500k license: apache-2.0 language: - en metrics: - accuracy --- # Qwen3-8B-ODA-Mixture-500k Leaderboard Performance Qwen3-8B-ODA-Mixture-500k is a supervised fine-tuned (SFT) model built on top of **Qwen3-8B-Base**, trained with **[ODA-Mixture-500k](https://huggingface.co/datasets/OpenDataArena/ODA-Mixture-500k)**. This training set is curated by mixing top-performing open corpora selected via the *[OpenDataArena](https://opendataarena.github.io)* leaderboard, and refined through deduplication and benchmark decontamination, aiming to improve the model’s general capabilities across **General**, **Math**, **Code**, and **Reasoning** domains. --- ## 🧠 Model Summary - **Base Model**: `Qwen/Qwen3-8B-Base` - **Training Data**: `OpenDataArena/ODA-Mixture-500k` - **Domain Coverage**: General, Math, Code, Reasoning - **Scale (selected training set)**: ~**500K** samples - **Goal**: Improve multi-domain reasoning and problem-solving ability while achieving stable overall performance gains. --- ## βš™οΈ Training Data Curation Pipeline ODA-Mixture-500k is built by following a single rule: **trust the OpenDataArena leaderboard**. ### 1️⃣ Data Collection We utilize **LIMO** as the foundational anchor due to its exceptional sample efficiency on the ODA overall leaderboard. To scale up to 500K, we integrate the top-performing and efficient corpora from specific ODA domain leaderboards, including: - **AM-Thinking-v1-Distilled-Math** (Math domain) - **AM-Thinking-v1-Distilled-code** (Code domain) - **math-gpt-4o-200k** (General domain) - **SYNTHETIC-2-SFT-verified** (Reasoning domain) ### 2️⃣ Deduplication & Decontamination We first perform **exact deduplication** over all questions to remove identical items, and then run **benchmark decontamination** to reduce evaluation leakage by removing overlaps with standard and competition benchmarks. ### 3️⃣ Data Selection At the 500K scale, our priority is distributional coverage. We employ semantic clustering to partition the total data pool into distinct thematic clusters. Within each cluster, we perform uniform sampling to ensure the final mixture represents a broad and balanced spectrum of reasoning tasks, maximizing the model's generalization capabilities. --- ## πŸ“š Training Data Source Composition | Source | Count | Percentage | |---|---:|---:| | LIMO | 817 | 0.16% | | AM-Thinking-Distilled-math | 150,244 | 29.67% | | AM-Thinking-Distilled-code| 150,252 | 29.67% | | math-gpt-4o-200k| 100,138 | 19.78% | | SYNTHETIC-2-SFT-verified| 104,913 | 20.72% | --- ## 🧩 Data Format The training data sample format is as follows (aligned with the dataset schema): ```json { "id": "unique_identifier", "source": "data source", "question": "textual question or instruction", "response": "textual response" } ``` --- ## πŸ“ˆ Performance Qwen3-8B-ODA-Mixture-500k is evaluated as an SFT model built on **Qwen3-8B-Base** across the full ODA benchmark suite spanning four domains: - **General (DROP, IFEVAL, AGIEVAL, MMLU-Pro)** - **Math (GSM8K, MATH500, Omni-Math, OlympiadBench, AIME2024)** - **Code (HumanEval, MBPP, LCB (V5), HumanEval+)** - **Reasoning (ARC-C, BBH, CALM, KOR-BENCH)**. We observe consistent improvements over the base checkpoint, with particularly strong gains on several benchmarks.
Leaderboard Performance Comparison. Best scores in bold, second-best underlined. Eff. denotes Data Efficiency.
Model / Training Data Size Eff. General Math Code Reasoning AVG
Qwen3-8B-Base
Qwen3-8B-Base -- 58.751.252.450.6 53.2
MiroMind-M1-SFT-719K 719k+0.023 64.577.263.665.8 67.8
AM-Thinking-v1-Distilled-math 558k+0.028 65.979.759.563.2 67.1
OmniThought-0528 365k+0.043 55.878.368.166.0 67.0
AM-Thinking-v1-Distilled-code 324k+0.045 64.864.975.859.3 66.2
Light-R1-SFTData 79k+0.168 64.971.859.063.6 64.8
SYNTHETIC-2-SFT-verified 105k+0.107 59.575.456.166.6 64.4
LIMO 817+0.490 61.746.052.754.1 53.6
ODA-Mixture-500k 500k+0.042 71.277.273.069.7 72.8
ODA-Mixture-101k 101k+0.177 61.177.373.264.7 69.0
--- ## 🌐 About OpenDataArena [OpenDataArena](https://opendataarena.github.io/) is an open research platform dedicated to **discovering, evaluating, and advancing high-quality datasets for AI post-training**. It provides a transparent, data-centric ecosystem to support reproducible dataset evaluation and sharing. **Key Features:** - πŸ† **Dataset Leaderboard** β€” helps researchers identify **the most valuable and high-quality datasets across different domains** - πŸ“Š **Detailed Evaluation Scores** β€” provides **comprehensive metrics** to assess data quality, complexity, difficulty, etc. - 🧰 **Data Processing Toolkit** β€” [OpenDataArena-Tool](https://github.com/OpenDataArena/OpenDataArena-Tool) offers an open-source pipeline for dataset curation and scoring. --- ## πŸš€ Usage Model repo: `OpenDataArena/Qwen3-8B-ODA-Mixture-500k`. Below is a minimal runnable example for loading and inference: ```python from transformers import AutoModelForCausalLM, AutoTokenizer MODEL_ID = "OpenDataArena/Qwen3-8B-ODA-Mixture-500k" tokenizer = AutoTokenizer.from_pretrained(MODEL_ID, trust_remote_code=True) model = AutoModelForCausalLM.from_pretrained(MODEL_ID, device_map="auto", trust_remote_code=True) messages = [ {"role": "user", "content": "Natalia sold clips to 48 of her friends in April, and then she sold half as many clips in May. How many clips did Natalia sell altogether in April and May?"}, ] text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True) inputs = tokenizer([text], return_tensors="pt").to(model.device) outputs = model.generate( **inputs, max_new_tokens=256, do_sample=True, temperature=0.7, top_p=0.9, ) print(tokenizer.decode(outputs[0], skip_special_tokens=True)) ``` --- ## πŸ“š Citation If you use this model or its training data (ODA-Mixture-500k), please cite: ```bibtex @article{gao2025closing, title={Closing the Data Loop: Using OpenDataArena to Engineer Superior Training Datasets}, author={Gao, Xin and Wang, Xiaoyang and Zhu, Yun and Cai, Mengzhang and He, Conghui and Wu, Lijun}, journal={arXiv preprint arXiv:2601.09733}, year={2025} } ``` ```bibtex @article{cai2025opendataarena, title={OpenDataArena: A Fair and Open Arena for Benchmarking Post-Training Dataset Value}, author={Cai, Mengzhang and Gao, Xin and Li, Yu and Lin, Honglin and Liu, Zheng and Pan, Zhuoshi and Pei, Qizhi and Shang, Xiaoran and Sun, Mengyuan and Tang, Zinan and others}, journal={arXiv preprint arXiv:2512.14051}, year={2025} } ```