OsaurusAI/NuExtract3-JANG_6M
JANG_6M build of numind/NuExtract3 for Apple Silicon: structured extraction (typed JSON templates), Markdown conversion and document chat over text and images, with the native MTP draft head kept.
Bundle
| On-disk size | 4.39 GiB (one shard), bfloat16 |
| Full attention q/k/v/o | 8-bit affine, group 32, GPTQ |
| Gated-delta in_proj_qkv / in_proj_z / out_proj | 6-bit affine, group 32, GPTQ |
| Gated-delta decay and beta gates (in_proj_a / in_proj_b) | 8-bit, GPTQ |
| MLP gate / up / down | 6-bit affine, group 32, AWQ + GPTQ |
| Tied embedding / lm_head | 8-bit affine, group 32 |
MTP draft head (mtp.*) |
2-D weights 8-bit group 64; norms raw |
| Kept at full precision | vision tower, conv1d kernels, A_log, dt_bias, all norms |
| GPTQ | 248/248 trunk tensors: fixed bf16 storage grid, act-order, full-corpus Hessians, down_proj Hessian rebuilt through the quantized gate/up, best-of-RTN guard |
Calibration (371,576 tokens)
Prompts are built by NuExtract3's own template (mode, template, enable_thinking) and
completed by the model itself with the card's sampling (0.2, or 0.6 when thinking), so its JSON,
Markdown and reasoning outputs are in calibration. Documents: synthetic invoices, emails,
resumes, contracts, tickets, statements and more; German Wikipedia; technical and security texts;
rendered receipts, tables, charts, notes and app screenshots.
| task | tokens |
|---|---|
| structured extraction from text (typed JSON templates) | 95,346 |
| document questions and summaries | 60,591 |
| template generation | 26,280 |
| image to Markdown | 81,254 |
| document detection | 18,779 |
| structured extraction from images | 57,361 |
| general text | 31,965 |
Measured against the bf16 source
47 held-out prompts never seen in calibration (7,537 scored answer positions), plus 23 held-out extraction tasks answered greedily by each build and compared with bf16's own JSON.
| build | size | KL mean | top-1 | valid JSON | fields matching bf16 |
|---|---|---|---|---|---|
| this bundle | 4.39 GiB | 0.00027 | 99.77% | 23/23 | 87.6% |
| same layout, round-to-nearest | 4.39 GiB | 0.00056 | 99.63% | 23/23 | 88.1% |
| uniform 6-bit gs64 (mlx-vlm default recipe) | 3.81 GiB | 0.00095 | 99.54% | 23/23 | 81.5% |
| numind/NuExtract3-mlx-4bits (official) | 2.83 GiB | 0.01118 | 98.24% | 23/23 | 77.5% |
Use
Pass NuExtract's template variables as chat-template kwargs (template, mode,
instructions, enable_thinking). Non-thinking extraction at temperature 0.2; thinking at 0.6.
- Downloads last month
- 1
6-bit