---
license: apache-2.0
task_categories:
  - text-generation
language:
  - en
tags:
  - speculative-decoding
  - distillation
  - kimi-k3
  - off-policy
  - atom
  - code
size_categories:
  - 100K<n<1M
---

# ATOM regen — OpenCodeInstruct, answered by Kimi-K3

Prompts from [`nvidia/OpenCodeInstruct`](https://huggingface.co/datasets/nvidia/OpenCodeInstruct)
with **every answer regenerated by Kimi-K3 itself**, at a **32768** window, for
off-policy speculative-decoding distillation.

**227,329 rows**, single-turn.

## Composition

Single source, single `source` value (`oci`). Selection:

| step | rows |
| --- | --- |
| rendered prompts, upstream | 4,825,044 |
| of those, tier-0 (upstream unit tests all pass) | 1,553,642 |
| sampled | 1,000,000 |
| generated and verified | 244,695 |
| **kept** | **227,329** |

**Only tier-0 was used.** Rows whose upstream `average_test_score` is perfect are
the low-entropy, template-shaped workload a speculator accepts best, which is
what this data is for.

| | mean | p50 | p90 | p99 | max |
| --- | --- | --- | --- | --- | --- |
| prompt tokens | 303.4 | 285 | 416 | 601 | 1,702 |
| answer tokens | 3,231.5 | 1,450 | 8,233 | 24,452 | 32,461 |

734,612,605 answer tokens in total. 9.8% of rows exceed 8192 tokens, 3.5% exceed
16384.

## Why code

In our own corpus, code deduped down to 0.43% of the mix and SWE-bench Pro was
the second-weakest benchmark at 84.9% of the reference draft. Raising code to 20%
— mostly with this data — took it to 100.6%, the largest relative gain of the
thirteen benchmarks we track.

Short prompts with long answers also make this a cheap source of long-context
positions compared with regenerating long prompts.

## Format

```json
{"conversations": [
  {"role": "user", "content": "..."},
  {"role": "assistant", "reasoning_content": "...", "content": "..."}
], "source": "oci"}
```

`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`](https://huggingface.co/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 (**7.10%**), led by unclosed thinking blocks.

## Limitations

- Single-turn only.
- tier-0 only, so the distribution skews template-shaped. Harder workloads would
  need tier-1.
- Prompts are short (p50 285 tokens); the long-context value is in the answers,
  not the prompts.