Ling-Coder-DPO: 250k Samples for Code Model Preference Tuning

Ling-Coder-DPO is a subset of 250,000 samples used for Direct Preference Optimization (DPO) training of the Ling-Coder Lite model. The dataset was created by inclusionAI and last updated on Hugging Face on March 27, 2025. It is part of a larger collection that also includes a supervised fine-tuning (SFT) subset with over 5 million samples and a synthetic question-answering subset.

Use Cases

• Training code generation models via Direct Preference Optimization based on the described DPO data.
• Fine-tuning language models for programming tasks using the companion SFT dataset mentioned in the description.
• Augmenting training pipelines with synthetic programming questions and answers as referenced in the description.

Strengths

• Contains 250,000 samples specifically for Direct Preference Optimization training.
• Part of a larger suite including an SFT subset with over 5 million samples.
• Explicitly created for training a named model, Ling-Coder Lite, indicating a clear purpose.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count for the full Ling-Coder-DPO subset is known, but the size of the synthetic QA subset is not fully specified.
• Description metadata is limited; actual data quality requires manual inspection after download.

Provenance

Source

inclusionAI via Hugging Face

Collection Method

Likely created for training the Ling-Coder Lite model, potentially involving synthetic generation and curation.

Time Range

null

Freshness

Last updated 2025-03-27 12:39:34; freshness should be verified.

Geography

null