Model Card for DeepSeek-R1-Distill-Qwen-1.5B

Table of Contents

1. TL;DR
2. Model Details
3. Usage
   • Uses
   • Out-of-Scope Uses
4. Bias, Risks, and Limitations
5. Training Details
6. Evaluation
7. Ethics and Safety
8. Intended Usage and Limitations
9. Benefits
10. Citation

---

TL;DR

DeepSeek-R1-Distill-Qwen-1.5B is a reasoning-focused lightweight model developed by DeepSeek-AI. Built using advanced reinforcement learning techniques, this model demonstrates strong performance in tasks involving math, code, and reasoning. DeepSeek-R1-Distill-Qwen-1.5B combines compactness and reasoning power, making it accessible for practical use cases in text generation, question answering, and conversational AI.

---

Model Details

Model Information

• Model Type: Text-to-text, decoder-only large language model
• Parameters: 1.5B
• Tensor Type: BF16
• License: MIT
• Related Models: DeepSeek-R1-Zero, DeepSeek-R1-Distill-Qwen series
• Hardware Compatibility: Compatible with both consumer GPUs and server environments.

---

Usage

Uses

DeepSeek-R1-Distill-Qwen-1.5B is well-suited for:

• Reasoning Tasks: Solving complex problems via chain-of-thought generation.
• Text Generation: Content creation, conversational AI, and summarization.
• Code Understanding: Processing programming-related prompts and generating code.

Out-of-Scope Uses

• Tasks requiring high factual accuracy or nuanced ethical understanding.
• Deployment scenarios with minimal computational resources.

---

Bias, Risks, and Limitations

Ethical Considerations

The model may reflect biases in its training data and generate unintended or inappropriate outputs. Mitigation measures include prompt engineering and evaluation.

Known Limitations

• Data Bias: Risk of propagating training data biases.
• Complexity Handling: May struggle with ambiguous or highly open-ended prompts.
• Factual Accuracy: Responses may include inaccuracies.

---

Training Details

Training Pipeline

• Base Model: Developed via large-scale reinforcement learning without supervised fine-tuning (SFT).
• Reinforcement Learning: Two-stage pipeline optimizing for reasoning patterns and human-aligned preferences.
• Distillation: Smaller models are distilled from the larger DeepSeek-R1 model, retaining high performance.

Datasets

• Sourced from publicly available data, including math, reasoning, and code repositories.

Hardware and Software

• Hardware: Trained on large-scale distributed systems.
• Software: Optimized using state-of-the-art machine learning frameworks.

---

Evaluation

Benchmark Performance
DeepSeek-R1-Distill-Qwen-1.5B demonstrates exceptional results on reasoning and language benchmarks:

Benchmark	Metric	Performance (%)
AIME 2024	Pass@1	28.9
MATH-500	Pass@1	83.9
LiveCodeBench	Pass@1	16.9
CodeForces	Rating	954

Configuration Recommendations

• Temperature Range: 0.5–0.7 (0.6 recommended).
• System Prompts: Avoid system prompts; include instructions in user prompts.

---

Ethics and Safety

Evaluation Approach
DeepSeek models undergo structured testing for content safety, representational harms, and memorization risks.

Results
The model meets internal standards for safety and performance, but users must exercise caution for high-risk applications.

---

Intended Usage and Limitations

Intended Usage

• Advanced Reasoning and Problem Solving:

  • Ideal for tasks requiring logical reasoning and chain-of-thought processes, such as math problem-solving, puzzles, and verification of complex workflows.
  • Supports step-by-step reasoning for better clarity and accuracy.

• Text Generation and Summarization:

  • Capable of generating dynamic, creative text, including stories, poems, marketing materials, and technical content.
  • Provides concise and accurate summaries of long documents, articles, or datasets.

• Conversational AI and Dialogue Systems:

  • Enhances virtual assistants and chatbots with context-aware, multi-turn conversations.
  • Supports interactive dialogue for customer support, education, or entertainment purposes.

• Education and Learning Assistance:

  • Acts as a tutor for personalized learning, explaining complex concepts step-by-step.
  • Supports language learning through conversational practice and grammar correction.
  • Aids researchers by generating summaries, references, or insights for academic work.

• Code Generation and Understanding:

  • Generates functional code snippets based on user instructions.
  • Identifies and resolves errors in programming through debugging support.
  • Automates the creation of detailed documentation and inline comments for codebases.

• Research and Development:

  • Enables NLP researchers to fine-tune models and test hypotheses on reasoning capabilities.
  • Provides a robust baseline for benchmarking language understanding and reasoning tasks.
  • Assists in exploratory data analysis by summarizing and generating insights from complex datasets.

• Decision Support Systems:

  • Powers business intelligence tools by reasoning through reports and providing actionable insights.
  • Evaluates scenarios for planning and operational problem-solving.

• Distillation and Model Fine-Tuning:

  • Supports creating smaller, high-performing models through distillation.
  • Enhances performance in reasoning and language generation tasks in constrained environments.

• E-commerce and Product Recommendations:

  • Enhances customer experience with personalized recommendations by reasoning through user inputs.
  • Improves chatbot-driven e-commerce platforms with insightful product suggestions and queries.

Limitations

• Bias Propagation:

  • May generate responses reflecting biases in the training data, which could affect ethical or cultural sensitivity.

• Factual Reliability:

  • Responses may not always be factually accurate, especially in ambiguous or nuanced scenarios.
  • Careful validation is required for high-stakes applications.

• Complexity in Handling Open-Ended Queries:

  • Struggles with vague, poorly structured, or highly nuanced prompts, leading to irrelevant or incoherent outputs.

• Resource Requirements:

  • Deploying the model in real-time applications or at scale demands significant computational power and memory, potentially limiting accessibility for low-resource users.

• Repetition and Coherence Challenges:

  • Without careful tuning of temperature or sampling strategies, the model may produce repetitive or incoherent outputs in certain scenarios.

• Limited Multimodal Support:

  • Lacks native capabilities to process or integrate non-textual inputs like images or audio without additional pipelines.

• Sensitive Data Risks:

  • The model may inadvertently generate or reveal sensitive information if improperly configured or prompted.

• Dependence on Prompt Quality:

  • Performance heavily relies on clear, specific, and well-structured prompts to achieve optimal outputs.

---

Benefits

DeepSeek-R1-Distill-Qwen-1.5B is a compact, high-performance model that democratizes access to advanced reasoning capabilities, supporting both research and commercial applications.

---

Citation

If you use DeepSeek-R1-Distill-Qwen-1.5B in your research, please cite:

copy@misc{deepseekai2025deepseekr1incentivizingreasoningcapability,
  title={DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning},
  author={DeepSeek-AI et al.},
  year={2025},
  eprint={2501.12948},
  archivePrefix={arXiv},
  primaryClass={cs.CL},
  url={https://arxiv.org/abs/2501.12948}
}

---

Contact

For support, please reach out to: service@deepseek.com.