README.md

metadata

license: mit
task_categories:
  - text-generation
tags:
  - instruction-tuning
  - sft
  - synthetic

Day 3 instruction blend

A normalized multi-source instruction dataset built for the Day 3 'instruction tuning at scale' exercise of a post-training curriculum. Every row is in the OpenAI-messages format with a source column for per-source ablations. Built by day03_instruct/build_blend.py; converters live in common/format_convert.py.

• Blend size: 24000 rows (200 held out as test)
• Sampling seed: 42
• Length cap: 8000 total content characters per conversation

Sources

source	dataset	original format	weight	rows kept
alpaca	yahma/alpaca-cleaned	alpaca	1.0	8000
slimorca	Open-Orca/SlimOrca	sharegpt	1.0	8000
ultrachat	HuggingFaceH4/ultrachat_200k	messages	1.0	8000

Filtering

Rows were rejected for structural problems (wrong role order, missing or non-final assistant turn, empty content, unknown speakers) or for exceeding the length cap. Per-source rejection counts:

{
  "alpaca": {},
  "slimorca": {
    "length: over 8000 chars": 96
  },
  "ultrachat": {
    "length: over 8000 chars": 1902
  }
}

Licensing and provenance

• alpaca (yahma/alpaca-cleaned): CC BY 4.0; responses originally distilled from OpenAI models
• slimorca (Open-Orca/SlimOrca): MIT; GPT-4-generated responses (note OpenAI-terms provenance caveat)
• ultrachat (HuggingFaceH4/ultrachat_200k): MIT

Portions of this data were generated by large language models; check the upstream dataset cards and the relevant providers' terms before commercial use.