hn.today

I turned Jev into a (lousy) chatbot

github.com62 points17 comments
Screenshot of I turned Jev into a (lousy) chatbot

jevchat converts Jev, a decision-style model, into a character/ token-level chatbot by repeatedly asking Jev a single-choice question: given the user's prompt and the reply so far, which symbol comes next? Each step presents an alphabet of options plus STOP, Jev returns a probability distribution, and a sampler draws the next symbol to append. The tool exposes multiple alphabets (lower26, ascii, tokens, words1k, bpe2k/5k) and a live interactive display showing partial replies, generation rate, top-scored options, and cancellation behavior; it was developed as a Claude-accelerated experiment intended to be amusing rather than cost-efficient.

Several sampling strategies and presentation modes change behavior and performance: choice (single question), bisect (binary splits), buckets (parallel grouped questions for large vocabularies), and refine (buckets followed by winner rescoring) are available, plus beam search to keep multiple candidate replies. Presenting options as finished hypotheses instead of raw symbols yields the largest improvement - roughly tripling top-1 accuracy and doubling mass on the correct symbol for character alphabets - while using fewer input tokens. Commands, configuration knobs (temperature, ensemble, repetition-penalty, bisect-cutoff, refine parameters), automated benchmarks, and 158 offline tests (no API key needed) are provided for experimentation.

Read on github.com17 comments on Hacker News

Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.

More in Other

The daily digest

Today's best Hacker News stories, summarized and screenshotted, one email a day.