---
tags: [agentic-ptb, kimi, training-data]
---

# kimi-data

Training data built by the AgentPTB arm for cell **`kimi`** — kimi-code / kimi-k3 @ effort `high`.

This is the corpus the arm *itself* assembled during its 100-hour run: what it downloaded,
filtered, rewrote and mixed. It is the input side of the checkpoints published as
`agentic-ptb/kimi.h*`, and the companion to the run record in `agentic-ptb/kimi-record`.

| field | value |
|---|---|
| plot cell | `kimi` |
| driver | kimi-code / kimi-k3 |
| reasoning effort | `high` |
| total size | 0.75 GB |
| path in run | `/mnt/pvc/users/simon/agentptb/runs/d/workspace/data` |

## Largest components

| directory | size |
|---|---|
| `sft_v2` | 0.26 GB |
| `sft_v1_f15k` | 0.25 GB |
| `sft_v2_parquet` | 0.23 GB |
| `rft_v1_parquet` | 0.01 GB |

## A note on apparent secrets

Automated scanning flags credential-shaped strings (`AKIA…`, `sk-…`, `AIza…`) in these
corpora. They were checked: they are **synthetic fixtures inside the training tasks
themselves** — secret-scanning and config-hygiene exercises whose task text embeds fake keys
(`"secret_type": "api_key"`, `"matched_pattern": …`). No credential belonging to this project
appears in the data.

## Related

- `agentic-ptb/kimi-record` — driver trajectory, harness, configs, evals for this cell
- `agentic-ptb/kimi.h*` — the checkpoints trained on this data
- `agentic-ptb/INDEX` — manifest joining every checkpoint back to the sweep figures