Instructions to use arcadia-impact/dispatch-models with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- PEFT
How to use arcadia-impact/dispatch-models with PEFT:
Task type is invalid.
- Notebooks
- Google Colab
- Kaggle
Commit History
-
scores/: the midtrain vs SDF-late vs graft three-way
Verified signatureAssembles a three-way comparison of true midtraining, SDF on a late instruct-tuned checkpoint, and grafting. It was never run as one experiment; the three studies are comparable because they pin the same corpus release (5c6eb06e) and the same agreement AFT file (8f28a074). scores/three_way_midtrain_sdf_graft/ carries the comparison itself, with a matched/unmatched ledger that has to be read before quoting an ordering -- notably that the graft's document stage is a rank-32 LoRA while the other two are full-parameter. Held-out-clause separations are published here for the first time: the wave's own RESULTS.md tabulates trained clauses only. They change the picture -- on trained clauses the three legs land within ~0.2 of each other, on held-out clauses midtrain +0.650 vs SDF-late +0.376. Five of the six source studies live on unmerged branches (sid/dispatch-lora-grafting-v1, sid/dispatch-graft-dose-v1, exp/token-scaling-law, sid/cookedness-dispatch-v1, sid/cookedness-grafting-v1), and …7b4adf1