sense-xs
Fab7's decision model, which Weft runs on your machine. Nothing leaves the machine.
One frozen microsoft/deberta-v3-xsmall reads every row, and a small LoRA
adapter for each job turns it to that job. Each row is
[CLS] text [SEP] question [SEP], read as one logit (yes minus no).
stateadapter, for a harness's pane:- Screen: the agent's terminal state, from the screen alone, in 7 states (idle, working, after its Stop hook, waiting on a permission, a question, trust or sign-in, or stopped on an error). Weft weighs this against the harness's own hooks; the model never sees them.
- Transcript entry: the person's message, the agent's reply, a tool step that worked or failed, or internal text.
- Turn end: whether the agent's reply asks the person something.
askadapter, for an Ask:- its task, its effects (read, write, execute, beyond the machine) and its route (direct, plan, goal, review); the result follows from these;
- Practice rules: each rule's own text is the question, so a rule you
write is read the same way as a shipped one. A rule's past use (how
often it was kept when offered) can lean its probability, with the
weight in
manifest.json; - Plan: whether the Ask continues, implements or amends a plan.
The questions, calibration and text normalizing are data in
manifest.json (schema fab7.sense.xs/1, normalizer
fab7.sense.decision.normalize/2).
Files
| File | What |
|---|---|
sense-xs.int8.onnx |
The base, with int8 weights and fixed scales. Input input_ids and attention_mask (int64[B, L], L ≤ 256); output logit (float32[B]). Each adapter's tensors are inputs with defaults. |
adapters/state.onnx_adapter, adapters/ask.onnx_adapter |
One adapter each: rank 16 on every encoder linear layer, plus its head (about 1.5 million numbers) |
tokenizer.json |
DeBERTa-v3's tokenizer |
manifest.json |
Every adapter's heads: their questions, classes and calibration; the normalizer; file digests |
golden_vectors.jsonl |
Normalized texts, token ids, logits and answers to check a runtime against |
Run it with ONNX Runtime 1.20 or later: load each adapter once
(LoraAdapter), and name it in each run's options
(RunOptions.add_active_adapter). Set session.disable_quant_qdq = 1, so
each row's answer is its own.
Training
- The base:
microsoft/deberta-v3-xsmall, first taught entailment on the human-labelled, commercially usable pairs ofMoritzLaurer/dataset_train_nli, then frozen. state: screens and transcripts from sandboxed harness sessions on local models, labelled by rules and checked by readers and a reviewer.ask: Asks written for every rule of Fab7's practice packs, its use cases and near misses, and real Asks from Fab7's own use. Two frontier models labelled each one through their coding harnesses, and a reviewer judged where they disagreed.- Each adapter: 3 epochs on one GPU.
Performance (0.0.2, test split, texts training never saw)
| Adapter | Head | Score |
|---|---|---|
state |
Screen | accuracy 0.981 (1,344 screens), from the screen alone |
state |
Transcript entry | accuracy 1.000 (433) |
state |
Turn end | average precision 0.931 (51) |
ask |
Task | micro F1 0.882 at 0.5 (1,000 Asks) |
ask |
Effects | micro F1 0.983 at 0.5 |
ask |
Route | accuracy 0.944 |
ask |
Practice rules | mean average precision within the pack: autonomous trading 0.871, machine learning 0.830, software development 0.830 |
The test split changed from 0.0.1 (more screens and Asks, rules relabelled). Both releases read on the same rows:
- Better than 0.0.1: screens (0.981 against 0.974, far fewer unsure), the plan an Ask names (0.87 against 0.70), and the task of Fab7's own Asks (F1 0.72 against 0.62).
- Worse than 0.0.1: practice rules in software development (0.83 against 0.93) and machine learning (0.83 against 0.85). The rule head now ticks more rules than apply. The route of Fab7's own Asks is 0.79 against 0.88.
The ask adapter trails a full fine-tune of the same model on the same
data by about 0.02 on task and 0.03 to 0.04 on rule ranking, and is within
noise on effects and route.
On a CPU, on an Apple Silicon Mac: a screen read (7 rows) takes about 76 ms, an Ask's task, effects and route (17 rows) about 195 ms, and a 20-rule pack about 300 ms.