---
license: apache-2.0
task_categories:
- text-generation
language:
- en
tags:
- kittylm
- persona
- reasoning
- think
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_think_train.jsonl
  - split: validation
    path: raw_think_val.jsonl
---

# KittyLM reasoning data

Think-augmented twin of the frozen V3 set in [KittyLM/kittylm-data](https://huggingface.co/datasets/KittyLM/kittylm-data):
same 900 + 100 prompts and answers, but every assistant reply is prefixed with a
`<think>...</think>` block of **plain-English private reasoning** (what the user asks,
the correct fact/plan, how to frame it as a kitten). The visible answer after `</think>`
is the untouched V3 kitten response.

Prompts are FROZEN at V3: newer data iterations reduced model performance, so no new
pairs were added — only `<think>` prefixes on the proven set.

## Why

Reasoning-capable bases (Granite 4, MiniCPM5 hybrid, LFM2.5-2.6B) trained on think-free
data never learned the think/answer split: on Ollama the kitten answer landed inside
the `Thinking` block with an empty visible reply. Training on this set teaches
English-think / kitten-answer separation. Non-reasoning models (Qwen3 non-thinking
variants, Gemma3, Phi-4-mini, LFM2.5-1.2B) keep training on `kittylm-data`.

## Files

- `raw_think_train.jsonl` / `raw_think_val.jsonl` — 900 + 100 raw `{"prompt", "response"}`
  pairs with `<think>` prefixes (built from V3, validated: think present exactly once,
  no kitten markers inside think, prompts byte-identical — see `scripts/validate_think.py`)
- `train.jsonl` / `eval.jsonl` — ShareGPT `{"messages": [...]}` with the same
  system-prompt mix as V3 (~50% KittyLM prompt, ~25% none, ~25% neutral)
- `system_prompt.txt` / `generic_systems.txt` — same files as V3