Needle — Lara local router (needle-lara-es-v1)

Fine-tune of Cactus-Compute/needle (a 26M-param Simple Attention Network) on the Lara on-device routing dataset: 22 Spanish assistant tools (time, date, weather, calendar, navigation, radio, search, etc.) plus delete/negative hard negatives.

Status: checkpoint artifact, NOT yet deployable on-device

This repository holds the JAX/Flax training checkpoint and the dataset used to produce it. Two things are still required before the senior-app can route locally with this model:

1. More training data. The base architecture needs roughly 120 examples per tool (100 train / 10 val / 10 test). This checkpoint was trained on only 75 total examples (3 per tool), which overfits. On the held-out gate it scores action_accuracy = 0.09 and call_f1 = 0.0 — it emits no tool call for most inputs and falls back to the server LLM. delete_false_positives = 0 (the delete guard holds).
2. A Cactus mobile bundle. cactus-react-native loads a Cactus-format int8.zip via its native runtime, not a JAX .pkl. Converting this checkpoint to weights/needle-lara-es-v1-int8.zip requires the Cactus conversion pipeline (not part of the open cactus-compute/needle repo).

Contents

• checkpoints/needle-lara-es-v1.pkl — fine-tuned JAX checkpoint (bf16).
• tools.es.json — the 22-tool contract used for training/inference.
• train.jsonl, val.jsonl — Lara dataset ({query, tools, tool_call}).
• val.preds.jsonl — this checkpoint's predictions over the gate val set.

Reproduce

# in the cactus-compute/needle checkout
needle finetune needle_train.jsonl --epochs 8 --batch-size 8

needle_train.jsonl is the dataset converted to Needle's {query, tools, answers} format (answers = JSON-encoded tool-call list, "[]" for negatives).