lucsaint/Deepseek-V4-Reasoning-Code-2500

DeepSeek Reasoning and Code Distillation Dataset

This dataset contains synthetic instruction-response examples generated from coding, reasoning, and math prompts. It was generated with enforce_distillable_text enabled using DeepSeek V4 Pro and DeepSeek V4 Flash through OpenRouter. It is intended for experimentation with supervised fine-tuning, response-style distillation, reasoning-format analysis, and code-assistant behavior research.

The dataset file is:

train.csv

It contains 2,555 filtered rows. Rows with finish_reason=length and rows with blank finish_reason were removed, so every remaining row has:

finish_reason=stop

Dataset Summary

Each row contains one user prompt and one assistant response. Assistant responses may include a reasoning block in this format:

<think>
reasoning content
</think>

final answer

The dataset was generated in three batches:

The prompts cover:

• —coding implementation tasks
• —software architecture and debugging tasks
• —data structures and algorithms
• —reasoning and math problems
• —probability, combinatorics, algebra, geometry, optimization, and logic

Columns

Models

The dataset contains outputs from DeepSeek V4 Pro and DeepSeek V4 Flash routed through OpenRouter with enforce_distillable_text enabled. Model identifiers present in the final CSV include:

• —deepseek/deepseek-v4-pro-20260423
• —deepseek/deepseek-v4-flash-20260423
• —deepseek-v4-flash

All retained rows were generated with:

reasoning_effort=high

Loading

from datasets import load_dataset

dataset = load_dataset("Banaxi-Tech/Deepseek-V4-Reasoning-Code-2500")
train = dataset["train"]
print(train[0]["user"])
print(train[0]["assistant"])

For a common instruction-tuning format:

def to_messages(row):
    return {
        "messages": [
            {"role": "user", "content": row["user"]},
            {"role": "assistant", "content": row["assistant"]},
        ]
    }

messages_dataset = train.map(to_messages)

Filtering Performed

The merged CSV was filtered before publication:

• —removed rows where finish_reason was length
• —removed rows where finish_reason was blank
• —checked for obvious API/HTTP failure payloads in assistant responses
• —kept only rows where finish_reason=stop

Final row count after filtering:

2555

Intended Uses

This dataset may be useful for:

• —supervised fine-tuning experiments
• —code-assistant behavior experiments
• —reasoning-format and answer-structure analysis
• —evaluating response parsing for <think>...</think> style outputs
• —small-scale synthetic data experiments

Limitations

This is a synthetic dataset generated by language models. It may contain:

• —incorrect reasoning
• —flawed code
• —insecure code patterns
• —hallucinated APIs or library behavior
• —overconfident explanations
• —solutions that require review before use

Do not use generated code from this dataset in production without human review, testing, and security analysis.

Rights and Licensing Notes

This dataset is released under the Apache 2.0 license. The referenced service terms assign rights in generated outputs to the user and permit applying inputs and outputs to use cases including academic research, derivative product development, training other models, and model distillation, as long as the usage is legal and follows the applicable terms.

Because the dataset includes reasoning traces inside <think> tags, downstream users should decide whether to keep, remove, or transform those sections based on their training objective and policy requirements.