Pleias CTO Pierre-Carl Langlais launches Sillon, a 600-million-parameter reasoning model for Paris subway operator RATP

@Scaleway Ok so with popular demand.

We designed a synthetic environment to model users' message and distress signals. We describe our experimental methodology for specialize synthetic environment at scale in a paper accepted to ACL finding. https://arxiv.org/pdf/2604.18226v1

Also what motivated the fast training to deployment cycle: tiny SLMs trained on synthetic environment outperformed all closed source alternative on RATP internal benchmark, elaborated independently with subway travellers.

A significant challenge was to recover the diversity of unformal French through generalized backtranslation and incorporate the RATP’s internal analytical specialized frameworks inside the model's reasoning traces.

Out: Fable In: Parisian Subway

The model was trained on public compute at Jean Zay and deployed this week on @Scaleway after only three months of development. Very fittingly, a 600m cousin of Baguettotron (is it Banettotron?) is now bundled inside the French subway.

For more details, @ynckdrt11496 will present our @aclmeeting paper in San Diego.

@chrisalbon I saw Parisian Subway Operator open at the state fair a few summers ago

@sharkey what a band name right?

@Dorialexander Love seeing AI applied to real infrastructure! This validates the targeted approach we're exploring in the UAE. 600M parameters beating the giants? Efficiency wins. #AI #Innovation

@Dorialexander @Scaleway We need a picture of baguettes taking the subway now, obviously

@Dorialexander What are the perfs when you fine-tune a top open weight like a recent qwen 8B on the same data as the small one ?