model card
Browse files
README.md
ADDED
|
@@ -0,0 +1,56 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: apache-2.0
|
| 3 |
+
base_model: Qwen/Qwen3.8-27B
|
| 4 |
+
library_name: mlx
|
| 5 |
+
tags:
|
| 6 |
+
- mlx
|
| 7 |
+
- decision-model
|
| 8 |
+
- solomon
|
| 9 |
+
- customs
|
| 10 |
+
- hs-code
|
| 11 |
+
---
|
| 12 |
+
|
| 13 |
+
# Solomon (DoccyHealth) — full BF16 MLX package
|
| 14 |
+
|
| 15 |
+
A ready-to-load MLX build of [DoccyHealth/Solomon](https://huggingface.co/DoccyHealth/Solomon), the open-weights
|
| 16 |
+
decision model (LoRA adapter + trained decision heads on `Qwen/Qwen3.8-27B`), assembled with the release's own
|
| 17 |
+
converter so that nothing has to be re-downloaded or re-converted on Apple Silicon. Everything in this repo is
|
| 18 |
+
Apache-2.0: the Qwen3.8-27B backbone (Alibaba Cloud), Solomon's adapter and heads (DoccyHealth), and the
|
| 19 |
+
conversion. The upstream `LICENSE`, `NOTICE` and `MODIFICATIONS.md` are included verbatim.
|
| 20 |
+
|
| 21 |
+
## What is inside
|
| 22 |
+
|
| 23 |
+
| part | files | provenance |
|
| 24 |
+
|---|---|---|
|
| 25 |
+
| backbone, BF16 (norms promoted before the unit offset, FP32-sensitive parameters kept in FP32) | `backbone/` (18 safetensors shards + tokenizer/config) | `Qwen/Qwen3.8-27B` @ `1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0`, converted by `solomon-mlx` (converter sha256 `810ec77e6a1e…`) |
|
| 26 |
+
| Solomon LoRA adapter | `adapter.safetensors` | `DoccyHealth/Solomon` @ `ac4f9ccf087861451e47277607c59b2ac3782ce4`, sha256 `2addaf841ecc…` |
|
| 27 |
+
| Solomon decision heads | `heads.npz` | same release, sha256 `126a9b5487dc…` |
|
| 28 |
+
| binding | `binding.json` | every file's sha256, the base hashes, the converter identity and the runtime contract (`solomon-mlx-binding-v1`, profile `quality`) |
|
| 29 |
+
|
| 30 |
+
Total 55.6 GB. `python -m solomon_mlx_hub verify <dir>` re-checks every pin without allocating tensors.
|
| 31 |
+
|
| 32 |
+
## Load
|
| 33 |
+
|
| 34 |
+
```python
|
| 35 |
+
from solomon_mlx import Solomon # pip install from the DoccyHealth/Solomon repo, mlx/ package
|
| 36 |
+
model = Solomon.load("path/to/this/repo")
|
| 37 |
+
```
|
| 38 |
+
|
| 39 |
+
The MLX package needs Apple Silicon and about 60 GB of unified memory at the `quality` (BF16) profile. The
|
| 40 |
+
`choice` question type takes 2–8 options; label questions take up to 64 candidates with one independent
|
| 41 |
+
probability each.
|
| 42 |
+
|
| 43 |
+
## Provenance note
|
| 44 |
+
|
| 45 |
+
Solomon's hub loader (`solomon_mlx_hub.prepare`) checks the licence texts of the release against hashes recorded
|
| 46 |
+
in its `release.json`. At revision `ac4f9ccf` the upstream `mlx/bf16/NOTICE` and `MODIFICATIONS.md` were edited
|
| 47 |
+
(commit `5c0a4a8`) without `release.json` being refreshed, so the check cannot pass on any copy of the repository.
|
| 48 |
+
This build tolerated that mismatch on those two licence texts only (warning emitted, patch documented in the
|
| 49 |
+
building project); the adapter, heads, base-manifest, converter-identity and every backbone checksum were
|
| 50 |
+
enforced as released. The `binding.json` here records the actual sha256 of the two texts.
|
| 51 |
+
|
| 52 |
+
## Why this exists
|
| 53 |
+
|
| 54 |
+
Built for the HS-code classification benchmark in `FrancoisChastel/HSCodeAgent` (decision-model arm next to
|
| 55 |
+
TypeSafe's Jev), where Solomon adjudicates a retrieved pool of customs precedents. Results and the evaluation
|
| 56 |
+
key are in that repository.
|