jackrong/dasd-4b-thinking-2507-stage2-gguf overview

{
  "metadata": {},
  "card_data": {
    "language": [
      "en",
      "zh"
    ],
    "license": "apache-2.0",
    "base_model": "Qwen/Qwen3-4B-Thinking-2507",
    "tags": [
      "qwen3",
      "unsloth",
      "text-generation",
      "reasoning",
      "math",
      "grpo",
      "sft",
      "distillation",
      "conversational"
    ],
    "pipeline_tag": "text-generation",
    "frontmatter": {
      "language": [
        "en",
        "zh"
      ],
      "license": "apache-2.0",
      "base_model": "Qwen/Qwen3-4B-Thinking-2507",
      "tags": [
        "qwen3",
        "unsloth",
        "text-generation",
        "reasoning",
        "math",
        "grpo",
        "sft",
        "distillation",
        "conversational"
      ],
      "pipeline_tag": "text-generation"
    },
    "hero_image_url": "",
    "summary": "**DASD-4B-Thinking-2507-stage2** is the final model in a three-stage training pipeline built upon Qwen/Qwen3-4B-Thinking-2507. It combines Reinforcement Learning via GRPO with a two-stage Supervised Fine-Tuning (SFT) strategy inspired by the **Distribution-Aligned Sequence Distillation (DASD)** methodology introduced by Alibaba Cloud Apsara Lab, resulting in a compact 4B model with enhanced mathematical reasoning and long chain-of-thought capabilities. ---",
    "quick_links": [],
    "benchmark_table_html": "",
    "readme_markdown": "---\nlanguage:\n- en\n- zh\nlicense: apache-2.0\nbase_model: Qwen/Qwen3-4B-Thinking-2507\ntags:\n- qwen3\n- unsloth\n- text-generation\n- reasoning\n- math\n- grpo\n- sft\n- distillation\n- conversational\npipeline_tag: text-generation\n---\n\n# DASD-4B-Thinking-2507-stage2\n\n**DASD-4B-Thinking-2507-stage2** is the final model in a three-stage training pipeline built upon [`Qwen/Qwen3-4B-Thinking-2507`](https://huggingface.co/Qwen/Qwen3-4B-Thinking-2507). It combines Reinforcement Learning via GRPO with a two-stage Supervised Fine-Tuning (SFT) strategy inspired by the **Distribution-Aligned Sequence Distillation (DASD)** methodology introduced by Alibaba Cloud Apsara Lab, resulting in a compact 4B model with enhanced mathematical reasoning and long chain-of-thought capabilities.\n\n---\n\n## 🧬 Training Pipeline Overview\n\nThis model is the culmination of three sequential training stages:\n\n```\nQwen/Qwen3-4B-Thinking-2507\n         │\n         ▼  Stage 0: GRPO (RL on Math & Reasoning)\nDASD-4B-Thinking-2507-GRPO-v2\n         │\n         ▼  Stage 1: SFT with Low-Temperature (T=0.6) Distillation Data\nDASD-4B-Thinking-2507-stage1\n         │\n         ▼  Stage 2: SFT with Default-Temperature (T=1.0) Distillation Data\nDASD-4B-Thinking-2507-stage2  ← (this model)\n```\n\n---\n\n## 📚 Stage Details\n\n### Stage 0 — GRPO Reinforcement Learning: `DASD-4B-Thinking-2507-GRPO-v2`\n\nStarting from the base model `Qwen/Qwen3-4B-Thinking-2507`, Group Relative Policy Optimization (GRPO) was applied using a high-quality mathematical reasoning dataset distilled from **DeepSeek-R1**. This stage significantly improved the model's:\n\n- Correctness on math problem solving\n- Step-by-step logical reasoning\n- Reward signal alignment for verifiable tasks\n\n---\n\n### Stage 1 — Low-Temperature SFT: `DASD-4B-Thinking-2507-stage1`\n\nInspired by the **Distribution-Aligned Sequence Distillation (DASD)** pipeline from [Alibaba-Apsara](https://huggingface.co/Alibaba-Apsara), Stage 1 SFT was performed using the **low-temperature subset (T=0.6)** of the [`Alibaba-Apsara/Superior-Reasoning-SFT-gpt-oss-120b`](https://huggingface.co/datasets/Alibaba-Apsara/Superior-Reasoning-SFT-gpt-oss-120b) dataset.\n\n#### 💡 Why Low-Temperature Distillation for Small Models?\n\nLow-temperature sampling from the teacher model (gpt-oss-120b) produces **sharper, more deterministic output distributions**, which are significantly easier for small student models to imitate and internalize. This \"cold-start\" strategy:\n\n- **Reduces distributional mismatch** between teacher and student — the cleaner, more peaked distributions generated at low temperature align better with what a small model can currently express\n- **Provides a stable foundation** — the model first learns the most consistent and representative reasoning patterns before being exposed to more diverse trajectories\n- **Boosts early performance rapidly** — low-temperature data provides an efficient jump-start for math and scientific reasoning benchmarks\n- **Mitigates exposure bias** — by gradually introducing complexity, the model avoids overfitting to noisy or outlier reasoning traces\n\nThis is the key insight behind DASD's temperature-scheduled learning: *start cold for stability, then warm up for diversity.*\n\n**Dataset used:**\n- [`Alibaba-Apsara/Superior-Reasoning-SFT-gpt-oss-120b`](https://huggingface.co/datasets/Alibaba-Apsara/Superior-Reasoning-SFT-gpt-oss-120b) — Stage 1 split (low-temperature, T=0.6)\n\n---\n\n### Stage 2 — Default-Temperature SFT: `DASD-4B-Thinking-2507-stage2` (this model)\n\nBuilding on `DASD-4B-Thinking-2507-stage1`, Stage 2 SFT was performed using the **default-temperature subset (T=1.0)** of the same dataset. Higher-temperature data introduces greater **lexical diversity and broader mode coverage**, enabling the model to generalize better across diverse reasoning patterns and problem domains.\n\n**Dataset used:**\n- [`Alibaba-Apsara/Superior-Reasoning-SFT-gpt-oss-120b`](https://huggingface.co/datasets/Alibaba-Apsara/Superior-Reasoning-SFT-gpt-oss-120b) — Stage 2 split (default temperature, T=1.0)\n\n---\n\n## 🗂️ All Datasets Used\n\n| Stage | Dataset | Purpose |\n|-------|---------|---------|\n| GRPO (RL) | [`a-m-team/AM-DeepSeek-R1-Distilled-1.4M`](https://huggingface.co/datasets/a-m-team/AM-DeepSeek-R1-Distilled-1.4M) | Math & reasoning RL training via GRPO |\n| SFT Stage 1 | [`Alibaba-Apsara/Superior-Reasoning-SFT-gpt-oss-120b`](https://huggingface.co/datasets/Alibaba-Apsara/Superior-Reasoning-SFT-gpt-oss-120b) (Stage 1, T=0.6) | Low-temp distillation, stable cold-start |\n| SFT Stage 2 | [`Alibaba-Apsara/Superior-Reasoning-SFT-gpt-oss-120b`](https://huggingface.co/datasets/Alibaba-Apsara/Superior-Reasoning-SFT-gpt-oss-120b) (Stage 2, T=1.0) | High-temp distillation, diversity & generalization |\n\nThe `Superior-Reasoning-SFT-gpt-oss-120b` dataset itself is built from the following upstream question sources:\n\n- [`nvidia/AceReason-1.1-SFT`](https://huggingface.co/datasets/nvidia/AceReason-1.1-SFT)\n- [`nvidia/OpenCodeReasoning`](https://huggingface.co/datasets/nvidia/OpenCodeReasoning)\n- [`nvidia/OpenScienceReasoning-2`](https://huggingface.co/datasets/nvidia/OpenScienceReasoning-2)\n- [`a-m-team/AM-DeepSeek-R1-Distilled-1.4M`](https://huggingface.co/datasets/a-m-team/AM-DeepSeek-R1-Distilled-1.4M)\n\n---\n\n## 🏃 Quickstart\n\n```python\nfrom transformers import AutoModelForCausalLM, AutoTokenizer\n\nmodel_name = \"Jackrong/DASD-4B-Thinking-2507-stage2\"\ntokenizer = AutoTokenizer.from_pretrained(model_name)\nmodel = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype=\"auto\", device_map=\"auto\")\n\nmessages = [\n    {\"role\": \"user\", \"content\": \"Solve: find all real solutions to x^3 - 6x^2 + 11x - 6 = 0.\"}\n]\n\ntext = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)\ninputs = tokenizer([text], return_tensors=\"pt\").to(model.device)\noutputs = model.generate(**inputs, max_new_tokens=4096)\nresponse = tokenizer.decode(outputs[0][inputs.input_ids.shape[1]:], skip_special_tokens=True)\nprint(response)\n```\n\n> **Tip:** This model naturally generates `<think>...</think>` reasoning traces before the final answer. You can parse these to inspect the chain-of-thought.\n\n---\n\n## 📋 Model Details\n\n| Attribute | Value |\n|-----------|-------|\n| **Base Model** | `Qwen/Qwen3-4B-Thinking-2507` |\n| **Architecture** | Qwen3 (4B Dense) |\n| **License** | Apache 2.0 |\n| **Language(s)** | English, Chinese |\n| **Training Framework** | [Unsloth](https://github.com/unslothai/unsloth) + Hugging Face TRL |\n| **RL Algorithm** | GRPO (Group Relative Policy Optimization) |\n| **Fine-tuning Method** | SFT (Two-stage temperature-scheduled distillation) |\n| **Developed by** | Jackrong |\n\n---\n\n## ⚠️ Limitations & Intended Use\n\n- This model is intended for **research and educational purposes** related to reasoning and mathematical problem-solving.\n- While mathematical and logical reasoning capabilities have been enhanced, the model may still produce incorrect answers — always verify outputs on critical tasks.\n- The model inherits the capabilities and limitations of the underlying `Qwen3-4B-Thinking-2507` architecture.\n- Not intended for deployment in high-stakes applications without additional safety evaluation.\n\n---\n\n## 📎 Related Models\n\n| Model | Description |\n|-------|-------------|\n| [`Qwen/Qwen3-4B-Thinking-2507`](https://huggingface.co/Qwen/Qwen3-4B-Thinking-2507) | Base model |\n| [`Jackrong/DASD-4B-Thinking-2507-GRPO-v2`](https://huggingface.co/Jackrong/DASD-4B-Thinking-2507-GRPO-v2) | After GRPO RL training |\n| [`Jackrong/DASD-4B-Thinking-2507-stage1`](https://huggingface.co/Jackrong/DASD-4B-Thinking-2507-stage1) | After low-temperature SFT |\n| [`Jackrong/DASD-4B-Thinking-2507-stage2`](https://huggingface.co/Jackrong/DASD-4B-Thinking-2507-stage2) | **This model** — final stage |\n\n---\n\n## 🙏 Acknowledgements\n\n- [Alibaba Cloud Apsara Lab](https://huggingface.co/Alibaba-Apsara) for the DASD methodology and the `Superior-Reasoning-SFT-gpt-oss-120b` dataset\n- [AM-Team](https://huggingface.co/a-m-team) for the DeepSeek-R1 distilled dataset\n- [NVIDIA](https://huggingface.co/nvidia) for open reasoning datasets\n- [Unsloth](https://github.com/unslothai/unsloth) for efficient fine-tuning infrastructure\n- [Qwen Team](https://huggingface.co/Qwen) for the excellent base model\n",
    "related_quantizations": []
  },
  "tags": [
    "gguf",
    "qwen3",
    "unsloth",
    "text-generation",
    "reasoning",
    "math",
    "grpo",
    "sft",
    "distillation",
    "conversational",
    "en",
    "zh",
    "base_model:Qwen/Qwen3-4B-Thinking-2507",
    "base_model:quantized:Qwen/Qwen3-4B-Thinking-2507",
    "license:apache-2.0",
    "endpoints_compatible",
    "region:us"
  ],
  "likes": 0,
  "downloads": 86,
  "gated": false,
  "private": false,
  "last_modified": "2026-02-24T02:29:28.000Z",
  "created_at": "2026-02-14T05:42:58.000Z",
  "pipeline_tag": "text-generation",
  "library_name": ""
}

{
  "_id": "69900b6261b2cd203fc67c98",
  "id": "Jackrong/DASD-4B-Thinking-2507-stage2-GGUF",
  "modelId": "Jackrong/DASD-4B-Thinking-2507-stage2-GGUF",
  "sha": "db29ed9476f4a16cac058273aea2d0135a99ffb3",
  "createdAt": "2026-02-14T05:42:58.000Z",
  "lastModified": "2026-02-24T02:29:28.000Z",
  "author": "Jackrong",
  "downloads": 86,
  "likes": 0,
  "gated": false,
  "private": false,
  "pipeline_tag": "text-generation",
  "library_name": "",
  "siblings_count": 14
}