bwen-dataset

The training data behind benthecarman/bwen-14b: a personal voice + opinion finetune built from @benthecarman's own tweets. No synthetic / AI-written text — every completion is a real tweet, and every prompt was hand-written by the author.

Built with the open pipeline at https://github.com/benthecarman/bwen (see the methodology).

Configs

config	rows	fields	what it is
pairs	301	prompt, completion, subject	Hand-written instruction → the real tweet that answers it. The core instruction-tuning set.
voice	3,103	text	Raw tweets used as a "voice layer" (trained with no prompt) to reinforce style.
eval	20	prompt, reference, subject	Held-out hand-labeled pairs for base-vs-tuned evaluation.

subject is the auto-discovered theme the tweet was clustered into.

How it was made

Parse a Twitter/X archive → filter (drop RTs, links, non-English; clean URLs/mentions) → discover themes (embeddings + UMAP + clustering) → score and surface a balanced shortlist → hand-write a prompt for each shortlisted tweet → add a raw-tweet voice layer. The prompts are triggers (the situation/question that elicits the tweet); the tweets carry the voice and opinions. Full details in the process doc.

Notes & license

This is one person's public tweets, restructured for finetuning, published by the author for research and reproduction. Opinions expressed are the author's. Treat accordingly — it's intended to clone a specific person's voice. To build the equivalent from your own archive instead, run the pipeline.