run-saina/saina-helm-0.8b-MLX

Saina Helm 0.8B — MLX

Apple-silicon build of run-saina/saina-helm-0.8b, a decision model that answers typed questions with probabilities. No text is generated.

Try it in the browser: Saina Helm decision demo (no install).

FP16, 1.9 GB (the untied lm_head adds about 0.5 GB). Loads with stock mlx-lm (0.32 or later); no custom runner is needed.

How it works

Helm's probability head is written into the lm_head rows of its answer-code tokens (A…Z, then the single-token two-letter codes). The logits at those tokens are Helm's head logits, so a softmax over the offered codes gives Helm's single-choice probabilities, and a sigmoid per code gives multi-label probabilities. Parity against the native PyTorch FP32 model stays within 0.002 probability across the test cases. FP16 is used rather than BF16: MLX's BF16 forward pass drifted by up to 0.023.

Don't quantize this build: score.py reads the code rows of lm_head directly, and quantizing would also blur the head.

Usage

pip install mlx-lm
hf download run-saina/saina-helm-0.8b-MLX --local-dir saina-helm-mlx

import sys; sys.path.insert(0, 'saina-helm-mlx')
from mlx_lm import load
from score import score

model, tokenizer = load('saina-helm-mlx')
score(model, tokenizer,
      'I was charged twice for my subscription this month.',
      'Which team should handle this?',
      ['Billing', 'Technical support', 'Shipping and delivery'])
# -> [p_billing, p_technical, p_shipping], summing to 1

score(model, tokenizer, 'Refund me, the package never arrived.', 'Which apply?',
      ['Refund request', 'Delivery problem', 'Praise'], mode='multi_label')
# -> independent probabilities, one per option

score.py builds the exact prompt Helm was trained on. If you build prompts yourself, keep them byte-for-byte identical, or the probabilities drift.

Versions

Version	Date	Changes
v1.0	2026-10-07	First public release.

Each release is a git tag on this repo, and main is the newest. Retrains that only change the weights bump the minor version (v1.1). Changes that need callers to change something (labels, prompt format, architecture) bump the major (v2.0). Scores can shift between versions, so pin a tag if you've tuned thresholds.

hf download run-saina/saina-helm-0.8b-MLX --revision v1.0 --local-dir saina-helm-mlx

On Ollama, the same release is run-saina/saina-helm-0.8b:0.8b-v1.0.

Other builds: weights (Transformers), GGUF, Ollama.