editmgt-e3
Private checkpoints for the editMGT E3 quality experiment and its DiMO follow-up. Each folder name states the experiment and checkpoint step. Frozen base models and training data are not included.
Checkpoints
editmgt-e3-quality-epoch1-step6250/: E3 quality pipeline, epoch 1, step 6250 (200,000 sample exposures). Contains trained adapter and region-conditioning weights, complete E3 resume state, and identity metadata. Source commit36dff5ff9ad10cee3d3910eb980932ec0a251b62.editmgt-dimo-e3-step7500/: DiMO E3 follow-up, step 7500 (60,000 sample exposures, 0.3 epoch). Contains student/auxiliary/EMA weights, complete distributed resume state, config and identity metadata. It was initialized from the E3 epoch1 checkpoint above. The full DiMO run is still in progress; this is an intermediate checkpoint, not a final quality-validated result. Source commitc4f92a7c696150c40f07959c72aa435cb6f67624.
The checkpoint configuration and manifests bind resume to their original model, code, data and training setup. See each folder's README for details.
Dense FP32 E3 checkpoints
- editmgt-e3-dense-fp32-epoch1.5-step9375: Dense full-Transformer SFT, epoch1.5 / step9375 /300000 sample exposures; full FP32 weights plus strict-resume optimizer/scheduler/sampler/RNG state. Intermediate checkpoint of the5epoch run.
Dense E3 checkpoints
- editmgt-e3-dense-fp32-epoch2-step12500 — Dense FP32, epoch 2.0, step 12500 (400k samples).
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support