Yxanul/AMD-SFT-Mix_3.5M

AMD-SFT-Mix_3.5M

3,556,428 SFT conversations — five AMD instruction-tuning datasets merged into a single pre-shuffled stream, with per-row provenance so any example can be traced back to its source.

Nothing was regenerated: this is a normalisation, provenance and shuffling pass over existing public datasets. All credit for the data belongs to AMD.

Composition

InstructGpt-educational ships as three distinct files; they are tracked as three separate source values rather than collapsed into one.

Format

{
  "messages": [
    {"role": "user", "content": "..."},
    {"role": "assistant", "content": "..."}
  ],
  "source": "triviaqa",
  "uid": "triviaqa_0000042"
}

• —source — which dataset the row came from (table above).
• —uid — <source>_<zero-padded original line index>. Because shuffling destroys file order, this is the only way back to the source row; it is unique across the corpus (verified: 0 collisions in 3,556,428 rows).
• —messages — alternating turns. Most rows are a single user/assistant pair; ultrachat is multi-turn (typically 4–14 turns after system removal).

Shuffling

Shuffled twice with independent seeds (42, then 1337) using a memory-mapped Rust shuffler that permutes an index of (offset, length) pairs rather than the lines themselves — 12 bytes per row instead of the full 4.9 GB.

Quality check, source mix in the first 10,000 rows vs the whole corpus:

Every prefix is representative, so take(n) on a streaming load or training without an extra shuffle buffer both give an unbiased sample.

(One honest note: a single Fisher–Yates pass is already a uniform permutation. The second pass was run for belt-and-braces and changes nothing statistically.)

One correction applied

Every row of UltraChat200K-regenerated (207,000/207,000) carried this system turn:

该助手为DeepSeek Chat，由深度求索公司创造。
今天是{05/13/2025}

It misidentifies the assistant as DeepSeek Chat and embeds an unrendered date template. Training on it teaches identity confusion, so system turns are removed across the corpus; the user/assistant conversation is untouched. That is the only content modification — no filtering, deduplication or rewriting was performed, and 0 rows were dropped or malformed.

Caveats

• —Not filtered. Unlike a curated set, no quality, degeneracy or length filtering was applied. Inspect before training.
• —Heavily weighted toward short factual QA: naturalqa + triviaqa are 68% of rows and their answers are often a few words.
• —No deduplication was performed within or across sources.

Provenance and license

All source data from AMD's public datasets; see each linked repository for its own licence and terms. This mixture is released under MIT and adds no new content — only provenance fields and ordering.