📖 Introduction

DistilQwen2.5-R1 Series: Advanced Reasoning Models

Overview

As large language models (LLMs) evolve toward deep reasoning capabilities, deploying them in resource-constrained environments (e.g., mobile devices, edge computing) remains challenging. The DistilQwen2.5-R1 series addresses this by transferring reasoning capabilities from ultra-large models (e.g., DeepSeek-R1) to compact models through innovative distillation techniques, achieving high performance while reducing computational costs.

Key Innovations

1. Cognitive Trajectory Adaptation Framework

Challenge: Discrepancies in reasoning paths between large and small models (e.g., small models struggle to comprehend large models' high-level problem-solving logic)
Solutions:
- Phase 1: CoT Data Optimization
  - Difficulty grading of large model reasoning chains (simple/medium/hard) via LLM-as-a-Judge
  - Adaptive adjustments: Expand simple chains and simplify complex chains to create medium-difficulty datasets digestible by small models
- Phase 2: Preference Optimization
  - Generate contrastive data pairs containing correct/incorrect reasoning paths
  - Apply DPO algorithm with tailored configurations to enhance reasoning path discrimination

2. Performance Highlights

DistilQwen2.5-R1-7B outperforms comparable distilled models (e.g., OpenThinker-7B) across multiple benchmarks
Successfully transfers high-order reasoning patterns originally dependent on large model parameter scales

Technical Advantages

Dynamic data optimization eliminates cognitive trajectory discrepancies
Two-stage training balances reasoning accuracy and computational efficiency
Enables complex task reasoning in edge computing environments

🚀 Quick Start

Here provides a code snippet with apply_chat_template to show you how to load the tokenizer and model and how to generate contents.

from transformers import AutoModelForCausalLM, AutoTokenizer
device = "cuda" # the device to load the model onto

model = AutoModelForCausalLM.from_pretrained(
    "alibaba-pai/DistilQwen2.5-R1-32B",
    torch_dtype="auto",
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained("alibaba-pai/DistilQwen2.5-R1-32B")

prompt = "Give me a short introduction to large language model."
messages=[
    {"role": "system", "content": "Your role as an assistant involves thoroughly exploring questions through a systematic long thinking process before providing the final precise and accurate solutions. This requires engaging in a comprehensive cycle of analysis, summarizing, exploration, reassessment, reflection, backtracing, and iteration to develop well-considered thinking process. Please structure your response into two main sections: Thought and Solution. In the Thought section, detail your reasoning process using the specified format: <|begin_of_thought|> {thought with steps separated with '\n\n'} <|end_of_thought|> Each step should include detailed considerations such as analisying questions, summarizing relevant findings, brainstorming new ideas, verifying the accuracy of the current steps, refining any errors, and revisiting previous steps. In the Solution section, based on various attempts, explorations, and reflections from the Thought section, systematically present the final solution that you deem correct. The solution should remain a logical, accurate, concise expression style and detail necessary step needed to reach the conclusion, formatted as follows: <|begin_of_solution|> {final formatted, precise, and clear solution} <|end_of_solution|> Now, try to solve the following question through the above guidelines:"},
    {"role": "user", "content": prompt},
]
text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True
)
model_inputs = tokenizer([text], return_tensors="pt").to(device)

generated_ids = model.generate(
    model_inputs.input_ids,
    max_new_tokens=2048，
)
generated_ids = [
    output_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids)
]

response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]

🔍 Evaluation

We compared DistilQwen2.5-R1 series with leading reasoning models across four benchmarks:

7B Model Comparison

Model	Training Data Size	AIME2024	MATH-500	GPQA Diamond	LiveCodeBench V2
DeepSeek-R1-Distill-Qwen-7B	800k	55.5	92.8	49.1	-
Bespoke-Stratos-7B	17k	20.0	82.0	37.8	36.1
OpenThinker-7B	114k	31.3	83.0	42.4	39.9
DistilQwen2.5-R1-7B	105k	43.33	88.4	42.93	46.38

32B Model Comparison

Model	Training Data Size	AIME2024	MATH-500	GPQA Diamond	LiveCodeBench V2
DeepSeek-R1-Distill-Qwen-32B	800k	72.6	94.3	62.1	-
Sky-T1-32B-Preview	17k	43.3	86.4	56.8	-
OpenThinker-32B	114k	66.0	90.6	61.6	68.9
DistilQwen2.5-R1-32B	105k	70.0	93.8	62.12	65.95

Key highlights:

DistilQwen2.5-R1 models achieve superior performance while using 6.1× less training data than DeepSeek-R1-Distill-Qwen series
Maintains open-source training lineage using filtered OpenThoughts subsets
Leads in LiveCodeBench V2 among open-source trained models