Skip to content

credence:e4b-calibrated

5 TagsUpdated 7.5B params8192 context100+ languagesApache-2.0by Txoka

Separate zero-synthetic checkpoint with a validation-fitted temperature: stronger public calibration, 72.20 % on typed decisions.

multilingual7.5b
ollaya run credence:e4b-calibrated --preset triage "I was charged twice for my subscription this month and want a refund."

Details

  • weights7c557d65b742 · 8.7 GBgguf · Q8_0 · huggingface.co/Txoka/Credence-v1-Gemma4-E4B/resolve/7d5ffc8…/calibrated/model-Q8_0.gguf
  • decision82df1e25c766 · 2 KB{"engine": "llama", "family": "winnow", "layout": "winnow-v1", "gguf": {…}, …}
  • calibration3769c01c5e88 · 240 B{"temperature": [1.041, 1.041, 1.041]}
  • licensea5aa25916b90 · 11 KBWinnow RLCD decision-model project

Every layer is checked against its sha256 when it is pulled. Weights and tokenizers download from the model author's Hugging Face repository at a pinned commit; Ollaya never re-hosts them.

Credence readme and all models