DeepSeek V4 Flash Distilled Dataset: 5K Synthetic Reasoning Traces

5,099 unique synthetic reasoning traces designed to mirror the step-by-step thinking style of the DeepSeek V4 Flash model. Generated entirely with template-based parameterized generation without using LLM APIs, this dataset was created by WithinUsAI and last updated on 2026-05-25.

Use Cases

• Fine-tuning language models for chain-of-thought reasoning based on the described reasoning traces.
• Benchmarking model performance on structured mathematical problems based on the 'domain' and 'difficulty' fields.
• Studying synthetic data generation techniques for educational or logical reasoning tasks.

Strengths

• Contains 5,099 unique reasoning traces.
• Generated without LLM API calls, suggesting a controlled and potentially reproducible synthetic process.
• Includes structured metadata fields like 'domain', 'subdomain', and 'difficulty'.

Limitations

• Description metadata is limited; actual data quality requires manual inspection after download.
• Column-level documentation is absent; field semantics must be inferred after download.
• Row count beyond the stated 5,099 is unknown, which may limit suitability assessment.

Provenance

Source

WithinUsAI

Collection Method

Template-based parameterized generation.

Freshness

Last updated 2026-05-25 01:30:43.