DASD-1.7B-stage2

Overview

DASD-1.7B-stage2 is a small-scale long-reasoning instruction model trained using the DASD (Distribution-Aligned Sequence Distillation) methodology. It employs a Stage-wise (Stage 1 → Stage 2) distillation fine-tuning strategy, specifically designed to enhance the model's stability, generalization, and reasoning diversity in complex tasks.

This model uses Jackrong/DASD-1.7B-stage1 as its initial weights. Building upon the stable reasoning patterns established in Stage 1, it undergoes Stage 2 high-temperature distribution alignment training to further expand its reasoning coverage, allowing the small model to better approximate the reasoning distribution of larger teacher models.

Training Methodology

Figure 1: (a) The stage-wise DASD pipeline. (b) Comparison of reasoning trajectories between stages. (c) Theoretical visualization of distribution alignment.

---

Base Model

• Base model: Jackrong/DASD-1.7B-stage1
• Architecture: Decoder-only Transformer
• Context length: up to 16,384 tokens
• Quantization: 4-bit
• Chat template: Qwen3 Instruct format

---

Training Method

• Training type: Supervised Fine-Tuning
• Frameworks:
  • 🤗 Transformers
  • TRL
  • Unsloth (4-bit efficient training)
• Optimization:
  • Optimizer: adamw_8bit
  • Learning rate: 2e-5
  • Scheduler: Linear
  • Weight decay: 0.001
  • Warmup ratio: 0.05
• Epochs: 1

---

Training Method: DASD Stage-wise Distillation

This model follows a two-stage Distribution-Aligned Sequence Distillation (DASD) training recipe, designed specifically to enable small dense language models to acquire strong long chain-of-thought (Long-CoT) reasoning behavior from large teacher models.

Instead of performing a single-pass SFT on mixed-temperature data, DASD decomposes training into two complementary stages with different optimization roles.

---

Stage 1: Low-Temperature Alignment (Stability)

In Stage 1, the model is trained on low-temperature teacher outputs, which emphasize:

• Canonical and structured reasoning paths.
• Reduced randomness and lower variance.
• High signal-to-noise supervision.

This stage serves as a distributional anchor, enabling the model to:

• Learn how to reason correctly before learning how reasoning can vary.
• Establish stable internal representations for long chain-of-thought.
• Avoid early exposure bias and optimization instability.

For small models, this stage is especially critical to prevent brittle reasoning and mode collapse during early fine-tuning.

---

Stage 2: High-Temperature Expansion (Diversity)

Stage 2 continues training from the Stage 1 checkpoint using higher-temperature DASD samples, which introduce:

• Multiple valid reasoning trajectories.
• Broader coverage of the teacher’s output distribution.
• Increased linguistic and structural diversity.

By aligning the student model to a wider but still distribution-aware target, this stage allows the model to:

• Generalize beyond a single reasoning style.
• Become more robust to paraphrasing and alternative solution paths.
• Better approximate the full reasoning support of large teacher models.

---

Why Stage-wise DASD Works for Small Models

Compared to conventional single-stage SFT, the Stage 1 → Stage 2 DASD pipeline provides:

• Improved optimization stability through curriculum-style training.
• Higher reasoning coverage without sacrificing correctness.
• Reduced exposure bias, especially for long-context reasoning.
• Better performance–data efficiency tradeoff, which is critical at the 1–2B parameter scale.

This strategy enables small dense models to inherit strong long-CoT reasoning behavior while remaining compact and efficient.

---

Training Data

The model is trained on multi-source, high-quality reasoning instruction data, primarily distilled from large teacher models using the DASD methodology.

Data sources include:

• Alibaba Apsara Superior-Reasoning-SFT (gpt-oss-120b) datasets (Stage 1 & Stage 2).
• Reasoning-focused ShareGPT-style instruction data.
• Natural and multi-step reasoning instruction corpora.
• Chinese and English bilingual long-CoT distillation datasets.

All data are unified into a conversational format using the Qwen3 Instruct chat template, with loss applied only to assistant responses.

---

Training Details

• Max sequence length: 16,384
• Loss masking: Assistant-only loss (train_on_responses_only)
• Checkpointing: Last checkpoint retained
• Logging: Experiment tracking enabled during training

---

Intended Use

This model is intended for:

• Long chain-of-thought reasoning tasks.
• Math, logic, code, and scientific QA.
• Bilingual (Chinese / English) instruction following.
• Research and experimentation on small-model reasoning distillation.

---

Limitations

• The model has not undergone RLHF or safety-specific alignment.
• Outputs may contain incorrect or unverified reasoning steps (hallucinations).
• Not suitable for high-risk domains such as legal or medical advice.

---

License

This model follows the licenses of its base model and training datasets. Please refer to the original dataset repositories for detailed license information.

---

Acknowledgements

• Alibaba Apsara DASD / gpt-oss-120b project.
• Unsloth efficient fine-tuning framework.
• Hugging Face Transformers & TRL teams.

---