license: mit
base_model: Cactus-Compute/needle
library_name: jax
pipeline_tag: text-generation
tags:
- function-calling
- tool-use
- termux
- android
- on-device
- candle
- rust
- needle
language:
- en
- ar
needle-termux-sting 🪡
Needle (26M-parameter "Simple Attention Network" for single-shot function calling, by cactus-compute) finetuned on the Termux:API command set — the model behind sting, a pure-Rust CLI that gives Termux natural-language device control, fully offline.
"vibrate for 2 seconds"
→ [{"name":"termux_vibrate","arguments":{"duration_ms":2000}}]
What's different from base needle
- Finetuned decoder — 4,810 synthetic examples over 16 Termux:API tools + 14 generic tools (single-call, multi-call, missing-argument, and no-tool cases; EN + some Arabic values). Recipe and data generator: sting/finetune.
- Working retrieval head — the released base checkpoint ships its contrastive (retrieval) head as all zeros, and zero weights + ReLU is a gradient fixed point, so ordinary finetuning can never revive it. This checkpoint's head was re-initialized and then trained on frozen encoder features (softmax-over-tools). Retrieval over the 16-tool Termux pack: hit@3 = 99.2%, hit@6 = 100% (400 queries).
Eval (held-out test set, 300 examples, 30 tools)
| metric | base needle | this model |
|---|---|---|
| call_f1 (name+args exact) | 75.0% | 99.7% |
| name_f1 | 94.6% | 100.0% |
| exact_match | 72.7% | 99.7% |
| args_acc | 79.2% | 99.7% |
| parse_rate | 99.3% | 100.0% |
Held-out but same-distribution synthetic data — treat as an upper bound for wild phrasing. Methodology + per-tool tables: sting/EVAL.md.
Files
| file | format | use with |
|---|---|---|
needle_sting_final.pkl |
needle checkpoint (JAX/Flax, f16) | the official needle pipeline: needle run --checkpoint needle_sting_final.pkl --query "..." --tools '[...]' |
model.safetensors + config.json + tokenizer_spec.json |
f16 safetensors + JSON specs | sting's pure-Rust candle runtime |
Usage (Python / needle)
from needle import SimpleAttentionNetwork, load_checkpoint, generate, get_tokenizer
params, config = load_checkpoint("needle_sting_final.pkl")
model = SimpleAttentionNetwork(config)
result = generate(
model, params, get_tokenizer(),
query="read the gyroscope, 5 readings",
tools='[{"name":"termux_sensor","description":"Read values from a hardware sensor on the device.","parameters":{"sensor":{"type":"string","description":"Sensor name: accelerometer, gyroscope, light, proximity, pressure, magnetic_field or gravity.","required":true},"limit":{"type":"integer","description":"Number of readings to take.","required":false}}}]',
stream=False,
)
# [{"name":"termux_sensor","arguments":{"sensor":"gyroscope","limit":5}}]
Usage (Termux / sting)
pkg install rust git binutils termux-api
git clone https://github.com/abod707/sting
cd sting && ./scripts/termux-install.sh
sting "turn on the flashlight"
Scope & limitations
Single-shot function calling over a provided toolset. Not conversational, no
multi-step planning; underspecified requests ("set an alarm" with no time)
correctly return []. Custom tools work zero-shot via the generic schemas it
saw in training; for production use of your own tools, finetune with ~120
examples per tool (recipe in the sting repo).
Credits
Base model, architecture, and training pipeline: cactus-compute/needle (MIT). Finetune, retrieval-head fix, and Rust runtime: abod707 (MIT).