pillardash/milogs-action-engine-2.0

Milogs Action Engine 2.0

A LoRA fine-tune of Cactus-Compute/needle3 that turns a spoken transcript into structured Milogs actions: journal entries, finance transactions, plans, reminders and search.

This is the model the Flotxt action router loads on device. It runs on the libneedle engine published with Needle 3, and it replaces the earlier NeedleForCausalLM Milogs Transformer.

The weights

file	role	size	sha256
needle3-tuned.cact	the model. Base checkpoint with the trained adapter merged in and the confidence head preserved.	63,474,900	4dcd641d8086762da35e8921844207fca5f7a263167ab4f4bda2c7cf276d60b5
whistle.cact	speech to text, 7 languages. Same engine, different archive.	16,919,407	b6e02f048568ac5d01a2042556c658061e699acbc0aa2a1439f52f3d461dffeb
adapter.safetensors	the rank-16 LoRA adapter. Continue training from this file, not from the merged archive.	7,906,680	714c7df454e25e250ab44f43ea0d5f285939c50c462c8dfbe0239bbf7f9498b1

needle3-tuned.cact and whistle.cact are both loaded by one engine. A process loads the text model and the speech model, and needle_complete then transcribes a clip and answers against your tools in a single call.

Base and architecture

• Base: Cactus-Compute/needle3 at revision 93e88ede867e4311c39151285e74df513e4b835b
• Architecture: NeedleForToolCalling, 20 layers, 121,021,910 parameters
• Attention: GQA with causal conv taps, sliding window 1024
• MLP: Monarch Hadamard, with engram n-gram memory and multi-lane hyper-connections
• Format: .cact, Cactus Quants
• License: Apache 2.0, following the base model

Fine-tuning

• Method: LoRA on the frozen base at the full 20 layers
• Rank 16, alpha 32
• Learning rate 4e-5, batch size 2, sequence length 2048
• 471 steps, resumed from the step-300 checkpoint of an earlier run, about 4 hours
• Training set: 1,046 examples, 34 percent of them refusal negatives
• Held-out validation loss: 0.2901

The training contract defines the tool schemas the model was taught. A Milogs release that consumes this model must validate arguments against the same schema, because the model emits exactly the field names it was trained on.

The confidence head

Refusal is not in the language weights. For a negative the model still proposes a call, but at a confidence near 0.005, and the engine suppresses anything below the threshold. The confidence head is the entire gate.

needle build --lora deletes confidence_head before export, because local tuning never trains it, so an archive built that way reports confidence: null and ships every proposed call. This archive is built by merging the adapter into the checkpoint by hand and exporting the merged checkpoint directly, so the head survives and real confidence values are reported.

Measured behaviour

43 hand-written cases, graded on an exact match of the full call list, on the same runtime the app loads.

metric	Milogs Action Engine 2.0	untuned needle3
exact match	12/43	13/43
right tool set	31/43	29/43
false fires on non-commands	8	5
real commands that fired	4/25	2/25
negatives correctly refused	5/8	7/8

Read this plainly. The adapter is a modest step. On exact match it is level with the untuned base, it routes to the right tool a little more often and fires on twice as many real commands, and it fires wrongly more often. It is not a solved model, and it should be evaluated against the release it replaces rather than assumed better.

Engine

The engine is published per platform in Cactus-Compute/needle3. On Android the app links android-arm64/libneedle.a; on iOS, ios-arm64/libneedle.a. Nothing in this repository needs to be paired with a specific engine build beyond what Needle 3 itself requires.

Citation

@software{cactus_needle,
  title  = {Needle: small language models for tool calling on device},
  author = {Cactus Compute},
  url    = {https://github.com/cactus-compute/needle}
}