sjsim/lky-reasoning-recipe

LKY Reasoning Corpus (a recipe)

A recipe for a training corpus that teaches Lee Kuan Yew's reasoning style to a language model. The style has four parts: reframe the question, argue from first principles, use a historical analogy, and end with a blunt conclusion. This corpus trained sjsim/lky-qlora, a LoRA for Qwen3-14B.

This repo is a recipe, not the transcripts. It ships the parts that are this project's own work: about 2,900 generated interviewer questions, plus pointers to the source records. A one-command step then fetches the transcript text from the National Archives of Singapore and builds the full training data on your machine. No archive text is redistributed here.

Pipeline and full method: https://github.com/pixiiidust/lky-brain

Why a recipe

The transcripts are records held by the National Archives of Singapore (NAS). They are Singapore government material, subject to NAS's conditions of use. So this dataset ships only what the project created and lets you pull the text from NAS yourself. That is more defensible. It also fits the project's goal better, which is a reusable template for teaching any documented expert's reasoning.

Not affiliated with the estate of Lee Kuan Yew or the Government of Singapore. Model outputs trained on this are AI-generated imitations, not real statements.

What's in this repo

file	contents
questions.jsonl	About 2,900 generated interviewer questions, one per speech passage (written by claude-opus-4-8). Each row also points to its source record: {question, record_uuid, nas_url, date, title, chunk_index}. This is the project's own work.
records.jsonl	A list of the roughly 1,300 source records: {uuid, title, date, source, nas_url}. Public archive metadata only. Pointers, no transcript text.
hydrate.md	The one-command rebuild, using the open-source pipeline.

You can browse questions.jsonl on its own. Each row is a generated question next to a pointer to where its answer lives in NAS.

Rebuild the full training corpus

The rebuild reuses the public pipeline. It costs no API spend, because the generated questions are already here, so you skip the generation step.

git clone https://github.com/pixiiidust/lky-brain && cd lky-brain
uv sync && uv run playwright install chromium

# add this recipe's questions, then run the pipeline without the paid backfill:
cp /path/to/questions.jsonl data/questions.jsonl
uv run python -m lky.crawl
uv run python -m lky.download
uv run python -m lky.extract
uv run python -m lky.parse_turns
uv run python -m lky.filter
uv run python -m lky.chunk_speeches
uv run python -m lky.build_dataset      # joins the questions to the fetched text
# -> data/train.jsonl + data/eval.jsonl

Chunking is deterministic. The same source PDFs produce the same passage boundaries, so each chunk_index pointer rejoins its question to the exact passage it was written for. See hydrate.md for detail.

What you get after rebuilding

{"messages": [...]} chat JSONL, ready for axolotl, unsloth, or LLaMA-Factory. About 4,400 training rows (interview windows, speech Q&A, and about 10% generic instruction data) and 66 held-out interview rows. Each row has a dated system prompt so the era is selectable at inference.

Limitations

• This transfers a reasoning style. It is not a factual record of his views.
• The source is mostly English. Some records are translations of Hokkien, Malay, or Mandarin remarks. It also reflects what NAS has digitised.
• The Apache-2.0 license covers this repo's contents: the generated questions, the pointers, and the structure. The transcript text you fetch from NAS is not covered by it and stays subject to NAS's terms.