ZhuoBohang-AMD/Kimi-K3-Draft-mxfp4-training-data

Kimi-K3 DSpark Draft · MXFP4 — training data

Training data of ZhuoBohang-AMD/Kimi-K3-Draft-mxfp4, used unchanged for all three epochs (57,849 steps) of its from-scratch MXFP4 run. Every final assistant answer was generated by Kimi-K3 itself. This is the same file as the final-stage data of the BF16 DSpark draft.

file	rows	size	sha256
data/train.jsonl	2,468,462	24.4 GB	0b584f2bcc800142cff81bdaf02e22c3a048cf4e0ca6b0711aea73166bb82a2c

Sources

Every row comes from one of three datasets. Each takes prompts from public datasets and replaces every answer with a Kimi-K3 generation (revision a590ce0, ATOM on MI355X, greedy, verified token-for-token).

upstream dataset	prompts from	rows
slippedJim/ATOM-regen-nine-category-prompt	nvidia/Nemotron-Post-Training-Dataset-v2, CohereLabs/aya_dataset	1,941,977
slippedJim/ATOM_regen_seeklight_kimi_mtp	lightseekorg/kimi-mtp-dataset	299,156
slippedJim/ATOM_regen_nv_OpenCodeInstruct	nvidia/OpenCodeInstruct	227,329

Image-bearing rows were removed, multilingual rows thinned, and prompts deduplicated so that a complete 32768-window answer replaces an earlier truncated one. The result was shuffled with seed 20260913.

Composition

category	share
multilingual (it, fr, ja, de, es)	36.74%
chat	17.04%
STEM	14.14%
kimi-mtp	12.12%
code, incl. OpenCodeInstruct	10.24%
aya	5.87%
math	3.85%

The per-row category is in source; rows from the kimi-mtp main file have none.

Format

{"conversations": [
  {"role": "user", "content": "..."},
  {"role": "assistant", "reasoning_content": "...", "content": "..."}
], "source": "chat"}

The final assistant turn is the distillation target. reasoning_content is Kimi-K3's thinking channel.

Training use

With the Lumen-RL example at a 32768 window, 2,468,457 rows survive preprocessing. The last 128 are held out for eval, which gives 19,283 steps per epoch at global batch 128, or 57,849 for three epochs.

License

other. Rows keep their upstream licenses: Apache-2.0 for the three upstream datasets, aya and kimi-mtp, and CC-BY-4.0 for the Nemotron and OpenCodeInstruct prompts. Answers are Kimi-K3 output under the Kimi-K3 license.