Plumb: Ornith wrote the curriculum. Match won.
Collection
Self-proposed G702 training tasks lose alone, help as a supplement. Code: https://github.com/caiotheodoro/plumb • 4 items • Updated
plumb-handseeded
Control. 223 tasks matched to the eval mix, including PASS. Study: collection.
| Base | mlx-community/Qwen3-1.7B-4bit |
| Train | 190 / 8 epochs / val loss 0.486 |
| Gold | caiotheodoro/plumb train_handseeded |
| sw-recall | precision | exact | parse | |
|---|---|---|---|---|
| this | 0.318 [0.290, 0.347] | 0.308 [0.279, 0.337] | 0.178 | 1.000 |
| Ornith-only | 0.241 [0.214, 0.268] | 0.111 [0.098, 0.124] | 0.084 | 0.997 |
| blend | 0.334 [0.306, 0.363] | 0.374 [0.342, 0.406] | 0.228 | 1.000 |
from huggingface_hub import snapshot_download
from mlx_lm.lora import load
path = snapshot_download("caiotheodoro/plumb-handseeded")
model, tokenizer = load("mlx-community/Qwen3-1.7B-4bit", adapter_path=path)
Limits
A 1.7B text policy, not the 27B multimodal production model. It reads the pay application as structured text rather than a rendered scan, so OCR is out of scope. All data is synthetic: generated AIA G702/G703 pay applications, not real contractor filings. CIs are 95% bootstrap, 10,000 resamples, seed 11, over the same 1000-task seed-777 benchmark. The clean-protocol N=18 arms and the pow-* / leaked-anchor grows are a null result and are not published as weights.
Hardware compatibility
Log In to add your hardware
Quantized