---
license: apache-2.0
task_categories:
- text-generation
language:
- en
tags:
- kittylm
- persona
- style-transfer
- synthetic
configs:
- config_name: processed
  data_files:
  - split: train
    path: train.jsonl
  - split: validation
    path: eval.jsonl
- config_name: raw
  data_files:
  - split: train
    path: raw_train.jsonl
  - split: validation
    path: raw_val.jsonl
---

# KittyLM data

Synthetic kitten-speak SFT data for the KittyLM family: every assistant reply is factually correct first, kitten-flavored second
(onomatopoeia like mrrp/nya/prrr, at most one `*action*` per reply, occasional `:3`).

## Files

- `raw_train.jsonl` / `raw_val.jsonl` — 900 + 100 raw `{"prompt", "response"}` pairs (cleaned: generator tic `"you're welcome"` removed except after real thanks, see `scripts/clean_dataset.py`)
- `train.jsonl` / `eval.jsonl` — ShareGPT `{"messages": [...]}` for **prompt-free** training: ~50% rows carry the KittyLM system prompt, ~25% have NO system message, ~25% carry a neutral prompt from `generic_systems.txt` (built by `scripts/prepare_dataset.py`)
- `system_prompt.txt` — the KittyLM persona prompt; `generic_systems.txt` — the neutral pool
- `generation_prompt.txt` (in project `data/`, not this repo) — the original Kimi generation prompt

## Generation

Seed examples + category-balanced synthetic generation (40% chit-chat / 30% factual QA / 20% instructions / 10% refusals),
then audited for repeated-phrase contamination (top repeated 3-gram check).