slippedJim/ATOM_regen_seeklight_kimi_mtp

ATOM regen — seeklight kimi-mtp, answered by Kimi-K3

Prompts from lightseekorg/kimi-mtp-dataset with every answer regenerated by Kimi-K3 itself, for off-policy speculative-decoding distillation.

469,086 rows.

Files

file	rows	generation window
data/train.jsonl	450,625	16384
data/train-long32k.jsonl	18,461	32768

train-long32k.jsonl regenerates the rows whose first pass stopped at finish_reason == max_tokens; 52.3% of them exceed 8192 tokens. It is a separate file, and the loader merges it into the train split automatically.

Composition

The main file carries no source field, so the table below covers the long-sample file only; the main file is the upstream mix as published.

sub-dataset	rows (long32k)
perfectblend	8,248
smoltalk_cn	5,675
continual_tool_kimi	2,903
creative_writing	556
kimi_2000x	413
dcagent	392
creative_writing_reasoning	274

Shape of the main file: 322,985 single-turn / 127,640 multi-turn (28.3%), turns p50/p90/max 2 / 8 / 140.

Both files are shuffled (seed 20260815). The upstream set is concatenated by source, not shuffled — idx 0–275k averages 770 prompt tokens, 350k–400k is a block of 141, and everything past 425k is long multi-turn at ~3,700 — so any prefix slice of the raw order is unrepresentative.

/!\ Both files carry only conversations, no source column. They have to match: putting files with different schemas in one split makes load_dataset fail when it concatenates the Arrow tables.

Format

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

reasoning_content holds Kimi-K3's thinking channel, stored separately from content. Multi-turn rows keep the full assistant reasoning of earlier turns: K3 is trained that way, and stripping historical thinking leaves empty think blocks that do not match the context seen at training time.

Generation

Teacher	moonshotai/Kimi-K3, revision a590ce090cb049c93a33dfe8c208ec652aa20503
Engine	ATOM 0.1.6rc1.dev275+g83e71001e (commit 83e71001e94602ee8e1d04581809bded49c3f08b)
Image	rocm/atom-dev:nightly_202608141640
Image digest	sha256:2f8bd4206ad15d014ae48115eae1ee9f1db83781848a8542de7177cfbd4ac914
torch / HIP	2.13.0+rocm7.14.0 / 7.14.60850
Hardware	8 x AMD Instinct MI355X per node, tensor_parallel_size=8
Sampling	greedy (temperature=0), one sample per prompt
KV cache	fp8 (matches the serving stack this data is distilled for)
Speculative decoding	off

Verification

Every row was re-rendered and re-tokenized, then compared token-for-token against what the engine actually consumed and produced. Non-matching rows were dropped (22.0% for the long-sample pass, which runs higher than short answers because longer generations are likelier to be cut off at the end).

Limitations

• One sample per prompt, greedy.
• Not bitwise reproducible even at temperature=0: batched MoE inference is not.
• The upstream set ships image-text rows whose answers the model invented, never having received an image. 122,812 such rows (27.3%) are present in the main file and should be filtered out before training on it.