jean-is/kev-0.6b-reflexes-ONNX

Kev-0.6B, fine-tuned for Jean's reflexes

Kev-0.6B fine-tuned on the questions Jean asks it about each message and each answer, in 33 languages, then put into the same ONNX graphs as onnx-community/kev-0.6b-ONNX: onnx/model_q4.onnx (4-bit MatMulNBits, blocks of 32, for WebAssembly) and onnx/model_q4f16.onnx (the rest in fp16, for WebGPU). Inputs, outputs, delimiters and usage are those of the original ONNX export, whose card applies as it is: only the weights of the 196 linear layers and of the pointer head differ, and the q4 graph's word lookup, rebuilt from core ONNX operators (training/fix_wasm.py: the same 4-bit table, packed two values to a byte, the same numbers out) where the original uses GatherBlockQuantized, which ONNX Runtime's WebAssembly build doesn't implement: this q4 graph loads in a browser without WebGPU, the original doesn't.

The questions

	Type	Question	State
personal	noul	Does the message share personal, health, financial or confidential information?	the user's message
unhappy	noul	Is the user unhappy with the previous answer?	previous answer: and new message:
lasting	noul	Does the message state a lasting fact about the user, a preference or a correction?	the user's message
demand	score	How demanding is this request? (a quick chat or a simple fact / reading, writing or summarising / several steps of work or analysis / expert reasoning)	the user's message
capability	choice	Which of these would help with this message, if any? (files, compute, none; either of the first two may be absent)	the user's message
table	noul	Does this text compare numbers across several items?	an answer
email	noul	Is this a decision or announcement that needs to be passed on to other people?	an answer

Results

Each share is where Jean would act as the label says (a yes at 0.6; demand right when its likeliest level gives the right tier of brain).

	Kev-0.6B	this model
Jean's test set, 33 languages × 44 (machine-translated from English, never trained on), q4 on WebAssembly	66%	96%
personal / unhappy / lasting	55% / 55% / 57%	100% / 92% / 98%
lowest languages: Yoruba / Swahili / Hausa / Amharic	55% / 52% / 57% / 55%	80% / 82% / 84% / 89%
held-out development set (1,075 rows of the training data's kind)	59%	97%
Kev's own test partition, decision-v7 (1,176 records): accuracy	0.808	0.817
the same: Brier score / confident errors (≥ 0.9, wrong)	0.266 / 4.0%	0.289 / 7.4%

The q4f16 graph agrees with the q4 one to 0.001 on the questions above.

Training

• Data (training/data/<lang>.jsonl): 6,175 labelled requests, about 190 per language, written natively in each language by Claude (Anthropic) to training/SPEC.md: each label balanced, with hard negatives (sensitive topics disclosing nothing, one-off instructions that look like preferences, negative words that aren't about the answer); none within a 0.6 character-trigram Jaccard of a test example. Each capability request also asked with one option left out, as Jean asks when one capability is already on.
• Replay: 2,172 records of Kev's own decision-v7 calibration and development partitions, so the model keeps its general skill (Jean also asks it free-form yes-or-no questions).
• Recipe: kev.train from jaredpalmer/kev with --init_from jaredpalmer/kev-0.6b (its LoRA r16 and pointer head), learning rate 5e-5, 2 epochs, batch 8, no option augmentation, fp32 on Apple Silicon (MPS), 1 h 49 min. checkpoint/ holds the adapter and head, for --init_from.
• Export (training/to_onnx.py): the LoRA merged, each linear layer quantized with ONNX Runtime's own 4-bit quantizer, written over the original graphs' weights; done on Kev-0.6B itself, it gives the published q4 file byte for byte.
• training/fix_wasm.py then rebuilds the q4 graph's word lookup for WebAssembly; its logits equal the GatherBlockQuantized graph's exactly.
• training/score.py and training/build.py measure and build as above; training/train.log is the run.

Limits

• The test set is small (8 examples per question and language at most) and its wording, translated from English, may be easier than real messages.
• Yoruba, Swahili, Hausa and Amharic stay lowest; unhappy and capability are the hardest questions there.
• Out of Jean's questions it is a little more often confidently wrong than Kev-0.6B.

License

Apache-2.0, as Kev-0.6B and its base, Qwen3-0.6B-Base.