decider-1.7B-MLX-6bit

A small decision model built for the LLM wiki: a personal wiki that a language model writes and keeps up to date from your own notes, the pattern described by Andrej Karpathy in 2026. The notes stay the source of truth; the model compiles them into topic pages and checks them as they change.

Keeping such a wiki honest takes many small closed decisions, again and again, on every new note: do two notes contradict each other? Is this value decided, or only proposed? Does a later note update an earlier one, or do two sources disagree? Does this note belong on this page? Asking a generative model each time is slow and noisy. This model answers them in one forward pass, from the scores of the option tokens, with a calibrated probability, and leaves the generative model only the writing.

question	input	options
contradiction	two notes	yes / no: do they contradict each other about the same thing?
hypothesis	one sentence	A decided / B only a proposal, offer, idea or hypothesis
update	an earlier and a later sentence	A the later one updates the first / B two sources disagree
membership	a topic page and a note	yes / no: does the note belong on the page?

Notes may be in Italian, English, Spanish, French, German or Portuguese; the questions are always in English. It was made for the wordspace notes app, where it runs on the Mac.

Fine-tuned from Qwen3-1.7B-Base with LoRA, fused, and quantized to 6 bits for MLX. See NOTICE for the full list of changes.

Use

pip install mlx-lm
hf download wordspace/decider-1.7B-MLX-6bit --local-dir decider
cd decider && python example.py

The prompt must be built with protocol.py, character for character: the model was trained on exactly that text. Read the answer from the logits at the last position: for each option take the best of its token forms (yes/Yes/YES, A/ A…), then a softmax over the options divided by the temperature in calibration.json (2.25). The temperature changes the confidence, never the chosen option.

Results

Same cases for both models. Gold: 154 hand-written note pairs from the wordspace app's tests, in seven sets across six languages. Liquid bench: 138 questions covering all four tasks. None of these cases were used for training; generated training cases sharing any eight-word span with them were discarded.

	gold	Liquid bench	total
LiquidAI d1-3B, MLX 4 bit	139/154	131/138	270/292
this model	143/154	137/138	280/292

By question (gold and Liquid bench together): contradiction 196/208 (d1 187), hypothesis 36/36 (35), update 24/24 (24), membership 24/24 (24). On a separate hand-written probe of 630 minimal pairs: 613/630.

The difference from d1-3B is not statistically significant (exact McNemar test: this model right alone on 19 cases, d1 on 9, p = 0.087).

Inside the wiki's contradiction check, where code first compares the facts it can read, this model decides the remaining pairs and a generative model writes quotes and reasons: on the 154 gold cases, 2 false alarms instead of 7 with d1-3B, and a third fewer calls to the generative model (not significant either: p = 0.07).

On an Apple silicon Mac, inside the app: 0.14 s per decision, 1.5 GB of memory, 1.3 GB on disk.

Limits

• Weaker than d1-3B on contradictions between numbers and dates (39/43, d1 43/43) and on implicit contradictions that need world knowledge (6/11, d1 8/11).
• Trained only on these four questions and this prompt format; other questions or wordings are untested.
• Inputs longer than 1024 tokens were not seen in training.
• Quantizing below 6 bits loses accuracy on the gold set (4 bit: 267/292).

Training

35,400 synthetic questions generated from hand-written templates in six languages, with counter-examples for known shortcuts (same subject reworded, short notes, documents as sources, compatible pairs). Validation used templates the training never saw. LoRA rank 16 on all layers, AdamW with learning rate 1e-5, gradient clipping at norm 1, early stopping on the validation loss (best at step 1500 of 4425). The loss is computed on the option tokens, read the same way as at inference. No personal notes were used.