Superior-Reasoning-SFT-gpt-oss-120b

📣 News

• Our dataset ranked #1 on the Hugging Face Datasets Trending leaderboard from January 20 to January 30.

🚀 Overview

The Superior-Reasoning-SFT-gpt-oss-120b dataset is a high-quality, open-source collection containing 435K samples designed to democratize the training of high-performance Long Chain-of-Thought (Long-CoT) models. Unlike standard distilled datasets that rely on random sampling or heuristic filtering, Superior-Reasoning-SFT-gpt-oss-120b is constructed using a principled Distribution-Aligned Sequence Distillation pipeline.

This pipeline addresses key limitations in current reasoning distillation—specifically distributional mismatch and exposure bias, enabling small dense models (like DASD-4B-Thinking) to achieve State-of-the-Art performance among open-source models of comparable scale and outperforms significantly larger models.

Superior-Reasoning-SFT-gpt-oss-120b contains the Stage 1 and Stage 2 SFT training data for DASD-4B-Thinking and can be used to reproduce the performance of DASD-4B-Thinking.

Key Features

• Extreme Data Efficiency: Contains only 435K total samples, an order of magnitude smaller than comparable datasets, yet yields SOTA results.
• Diverse Domain Coverage: Spans Mathematics, Code Generation, Scientific Reasoning, and Instruction Following.
• Teacher Provenance: All data is distilled from gpt-oss-120b using its high‑reasoning mode, ensuring high-quality reasoning traces.

Training Tip

As detailed in the technical report, our training pipeline integrates three complementary mechanisms:

• Temperature-scheduled Learning: combines low-temperature trajectories (stability) with higher-temperature trajectories (diversity and exploration), improving both learning stability and coverage of teacher's modes.
• Divergence-aware Sampling: selects sequences based on teacher–student distributional mismatch at the sentence level and find target sequence-level distribution better supports effective learning. We also provide logprob of teacher models in our Superior-Reasoning-SFT-gpt-oss-120b-Logprob Dataset.
• Mixed-policy Distillation: introduces a lightweight on-policy revision protocol after off-policy SFT, using student-generated prefixes with teacher rewrites to improve robustness and mitigating exposure bias of distilled model

Together, these components form a coherent, distribution-aware sequence distillation pipeline that improves the fidelity of reasoning transfer while remaining data-efficient, making strong long-reasoning behavior feasible in a 4B dense model.

The Superior-Reasoning-SFT-gpt-oss-120b dataset comprises training samples used in Temperature-scheduled Learning, divided into two stages:

• Low-Temperature Training (Stage 1, ~105k)
• High-Temperature Training (Stage 2, ~330k).

(Note: The training data for the mixed policy distillation stage is generated by the model itself after completing Stage 1 and Stage 2 training, and is therefore not included in this dataset.)

All samples across both stages are generated via Divergence-aware Sampling and subjected to rigorous quality control measures, such as filtering truncated outputs and repetitive content.

For best results, we recommend a staged training recipe:

• Train on Stage 1 first for stability
• Then continue training on Stage 2 for diversity.

(Note: Our DASD-30B-A3B-Thinking-Preview model also achieved strong results using only Stage 1 data).

As an example, when training Qwen3-4B-Instruct-2507, we suggest following the configuration for Stage 1 and configuration for Stage 2.

Data Sources

We use the following publicly available datasets as question sources:

• nvidia/AceReason-1.1-SFT
• nvidia/OpenCodeReasoning
• nvidia/OpenScienceReasoning-2
• a-m-team/AM-DeepSeek-R1-Distilled-1.4M

📄 Data Format

The DASD-Thinking dataset adheres to a structured JSON format, with each instance containing the following fields:

Sample:

{
    "uuid": "a7c94c27-1c6a-434a-bec9-dd90c1acac28",
    "input": "Find the sum of the first six terms of an arithmetic sequence in which the sum of any number of terms is equal to four times the square of that number. Let's think step by step and output the final answer within \boxed{}.",
    "output": " We need to parse the problem: \"Find the sum of the first six terms of an arithmetic sequence in which the sum of any number of terms is equal to four times the square of that number.\" So we have an arithmetic sequence (i.e., terms are a, a+d, a+2d, ...)...",
    "domain": "math",
    "meta": {
      "training_stage": "stage1",
      "sampling_temperature": 0.6,
      "teacher_model": "gpt-oss-120b",
      "logprob_filename": "stage1/math/train-00073-of-00096.parquet"
    }
}

Dedicated Fields

• uuid (str): A unique identifier for the sample.
• input (str): The model input prompt, typically a query or instruction.
• output (str): The corresponding teacher-generated response.
• domain (str): The task domain (e.g., "math", "code", "science"), indicating the nature of the problem.
• meta (dict): A metadata dictionary containing auxiliary information used during data generation and training:
  • training_stage (str): The training stage the sample belongs to (e.g., stage1 for Stage 1, stage2 for Stage 2).
  • sampling_temperature (float): The temperature used by the teacher model when generating the output.
  • teacher_model (str): The name or identifier of the teacher model that produced the response.
  • logprob_filename (str): The filename containing the corresponding log probabilities in Superior-Reasoning-SFT-gpt-oss-120b-Logprob Dataset.

📊 Proven Effectiveness

Models trained on this specific dataset recipe achieve State-of-the-Art performance for their size class.

4B Dense Model Performance

In both Stage 1 and Stage 2, we use an initial learning rate of 5e-5 and a global batch size of 64, training for 6 epochs. Stage 2 continues training from the final checkpoint of Stage 1. Detailed training configurations can be found in our GitHub repository or technical report.

30B MoE Model Performance

DASD-30B-A3B-Thinking-Preview (trained on Stage 1 data only) demonstrates incredible data efficiency.

DASD-30B-A3B-Thinking-Preview has only undergone the first stage of training, using an initial learning rate of 5e-5 and a global batch size of 64 for 6 epochs. Detailed training configurations can be found in our GitHub repository or technical report.

📜 Dataset Access & License

The dataset is released under CC BY 4.0.

📚 Citation

DASD-Thinking is developed by Alibaba Cloud, as part of our mission to advance open, efficient, and trustworthy reasoning systems. If you find this work useful in your research or applications, please cite our technical report.

@article{yan2026dasd,
  title={Distribution-Aligned Sequence Distillation for Superior Long-CoT Reasoning},
  author={Yan, Shaotian and Liu, Kaiyuan and Shen, Chen and Wang, Bing and Fan, Sinan and Zhang, Jun and Wu, Yue and Wang, Zheng and Ye, Jieping},
  year={2026},
  journal={arXiv preprint arXiv:2601.09088},
  url={https://arxiv.org/abs/2601.09088}
} 
    
@article{liu2025where,
  title={Where Did This Sentence Come From? Tracing Provenance in LLM Reasoning Distillation},
  author={Liu, Kaiyuan and Yan, Shaotian and Miao, Rui and Wang, Bing and Shen, Chen and Zhang, Jun and Ye, Jieping},
  journal={arXiv preprint arXiv:2512.20908},
  year={2025}
}

We welcome collaboration, feedback, and community contributions to push the boundaries of what small models can reason about—transparently and responsibly.