Skip to content
Blame History Permalink
  • David FORDEN's avatar
    Fix TranslateGemma repetition loop: add repetition_penalty and cap max_tokens · d26eb7dc
    David FORDEN authored and Jérôme Revillard's avatar Jérôme Revillard committed 
    The completions API calls had max_tokens=4096 with no repetition_penalty,
causing TranslateGemma to degenerate into infinite phrase repetition
(e.g. "e mahere" repeated hundreds of times in Sesotho output).

- Add repetition_penalty=1.2 to all TranslateGemma completions payloads
- Reduce max_tokens from 4096 to 2048 (source text is typically <500 tokens)
- Apply same fix to both query translation and history translation paths
- Update CHANGELOG.md
    d26eb7dc
To find the state of this project's repository at the time of any of these versions, check out the tags.