Skip to content
Elon Salfati
Publications

ModelHugging Face

René-1: A 31B Decision Reader That Knows When It Is Sure

An open-weights 31B model that reads a document and returns a calibrated probability for every option of every question, in one pass.

Elon Salfati

View on Hugging Face

Key takeaways

  1. 01It answers typed questions about a document (yes or no, one of several options, or a point on a scale) with a probability for every option, and generates no text.
  2. 02It is a full fine-tune of a 31B open-weights model: one epoch on 396,500 rows from 56 task families, trained with a log-score loss to put probability on the right option.
  3. 03The model card reports 64.18 balanced skill on 37 of the 40 Decision Index 0.2 benchmarks, against 54.75 for the next model, and a median calibration error of 0.046. These figures are self-reported.
  4. 04Stored in FP8, it serves on one GPU: 92.1 ms at the median for five questions over 641 input tokens on one NVIDIA B200.
  5. 05The weights are open under Apache 2.0.

Abstract

René-1 is a decision model. You send a document and a set of typed questions: pick one option, answer yes or no, or place the document on a scale. It reads the document and every question in one pass, each question on its own, and answers each with a probability for every option you declared. There is no generated text to parse. The weights are stored in FP8, so the whole model runs on one GPU.