LosslessMegaCodeTrainingV3 Tiny: 650k Evol Instruction Lines for Code-Focused LLMs

A 2023 experimental dataset by rombodawg containing 650,000 instruction-formatted lines of data. The content is roughly 80% coding instruction data and 20% non-coding instruction data, intended to preserve logic and reasoning skills during model training. This is a refined version of the LosslessMegaCodeTraining series.

Use Cases

• Fine-tuning code generation models based on the 80% coding instruction data.
• Training instruction-following models using the evol instruction format mentioned in the description.
• Improving model reasoning skills by leveraging the 20% non-coding instruction data.
• Creating benchmarks for evaluating code-specific language model performance.

Strengths

• Contains 650,000 instruction-formatted lines of data.
• Explicitly designed with a 80/20 split between coding and non-coding instruction data to balance skills.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Last updated 2023-09-07 23:48:13; freshness should be verified.

Provenance

Source

huggingface user rombodawg

Collection Method

Likely aggregated and refined from various sources for instruction tuning.

Freshness

Last updated 2023-09-07 23:48:13.