Skip to content

largha

4 TagsUpdated Apache-2.0by Phocinae

Phocinae's Largha: Laya's decision network on mmBERT-small (144M), fine-tuned on the typed-decisions workflows. Small enough for a CPU, on the Apple GPU through MLX.

multilingualfine-tuned144m
ollaya run largha --preset triage "I was charged twice for my subscription this month and want a refund."

Models

View all
Name
larghalatest328 MB · 512 ctx · 100+ languages
largha:150m328 MB · 512 ctx · 100+ languages

Each model carries fp16 and fp32 graphs over one weights file, and loads fp16 on a CUDA GPU and fp32 on CPU.

Readme

Largha by Phocinae is a small decision model: Laya's decision network on mmBERT-small (144M parameters), fine-tuned on the typed-decisions workflows. It reads a state and typed questions and returns calibrated probabilities for every question in one forward pass. It never generates text.

Needs Ollaya 0.15.0 or newer.

Models

TagBackboneParamsContextLanguages
largha:latest, largha:150mmmBERT-small144M512100+ (the author evaluated English and Chinese)

largha:150m carries an fp16 and an fp32 graph over the author's own weights file. It loads fp16 on a CUDA GPU and fp32 on the CPU; largha:150m-fp16 and largha:150m-fp32 pin one. On Apple silicon it runs on the GPU through MLX.

Usage

ollaya run largha --preset triage "I was charged twice for my subscription this month. Please refund the second charge."

Point any TypeSafe client at http://localhost:11435 and set the model to largha.

Accuracy and speed

Measured by Ollaya on an RTX 5090: 0.398 macro accuracy on Bespoke Labs' public benchmark (3,880 human-labeled questions, pooled ECE 0.026), and 18 ms for five questions through the HTTP API (339 ms on the CPU of the same machine).

Largha has no typed-decisions score here. It was fine-tuned on that dataset's train split, and by its card its training also saw 500 of the 2,000 test decisions, so its typed-decisions score is not a held-out measurement. The author reports 0.5455 (126 of 231) on JevBench's public items.

How it works

  • Network. The same network as laya: the question, a [MASK] before every option, and the state go into one sequence, and a two-layer head scores each option at its [MASK]. The author's rl_agent_config.json sets the budget (512 tokens, 192 for the question and its options) and the three temperatures Ollaya applies.
  • Parity. Ollaya's runtime matches laya's own loader on the author's v1.1 checkpoint in fp32 on 2383 questions, on the CPU, on CUDA and on the Apple GPU: identical token ids and markers, the same decision on every question, probabilities within 3.5e-6. The fp16 graph agrees on 2380 of 2383 decisions; the others are near-ties.

Limits

  • State. 512 tokens including the question. /v1/systemone answers a longer state with STATE_TRUNCATED; /api/decide answers from its first tokens.
  • Options. Every option's [MASK] has to fit in the 512-token sequence; long options are shortened to share the 192-token question budget, and options that still don't fit return TOO_MANY_OPTIONS.
  • Calibration. Ollaya applies the author's three temperatures. The repository's calib/ power transform is not applied.

Weights and license

Apache-2.0. Largha is by Phocinae; its base encoder, jhu-clsp/mmBERT-small, is MIT. Ollaya downloads the weights from Phocinae/Phocinae-Largha-150M-v1, pinned to the v1.1 commit and checked against sha256; it never re-hosts them.