← Builder galleryReference pattern

Encoder on a frozen backbone

A small modality encoder projects structured events into a frozen LFM2.5-350M; LoRA plus task heads read out decisions.

4
internal demonstrations of the recipe
~2 MB
per personal/vertical adapter
10–50ms
warm inference, internal measurements
Demo-grade — internal demonstration

The four encoder builds behind this pattern are internal demonstrations of the recipe, trained and evaluated on synthetic data. They demonstrate feasibility on this architecture — treat every number as an internal measurement on synthetic distributions.

The recipe

Tabular or event-sequence data (transactions, telematics, claims, security telemetry) is encoded per-field, projected into the backbone's embedding space via inputs_embeds, and decisions are read out by small task heads. The backbone stays frozen; the vertical lives entirely in the encoder + LoRA + heads.

The same shape generalizes across verticals — payments, mobility, claims, and cyber were built from one template with different encoders and head sets.

64 events × features → field embeddings → encoder
  → inputs_embeds → LFM2.5-350M (frozen) + LoRA
  → heads: fraud · next-event · risk · …

Recorded runs

This pattern page ships with recorded end-to-end runs rather than public live instances — the live builds sit behind the gated tier where an engineer can frame them properly.

Live examples of this pattern

Start from code