TTangenty commited on
Commit
a772800
Β·
verified Β·
1 Parent(s): c87bf7e

Organize E3 and DiMO checkpoints into named folders

Browse files
README.md CHANGED
@@ -5,28 +5,13 @@ tags:
5
  - lora
6
  - checkpoint
7
  ---
8
- # editMGT E3 epoch 1 (step 6250)
9
 
10
- Complete checkpoint from the E3 ROI-conditioning experiment, frozen D200K corpus, 8 GPUs, global batch 32. Step 6250 represents 200,000 sample exposures (1 epoch). This is the E3 teacher used to initialize the subsequent DiMO experiment, not a DiMO student checkpoint.
11
 
12
- Source code: https://github.com/yiyezhiqiu2077/editMGT/tree/fix/e3-quality-pipeline-v1
13
 
14
- Source commit: `36dff5ff9ad10cee3d3910eb980932ec0a251b62`.
 
15
 
16
- ## Contents
17
-
18
- - `adapter_model.safetensors`: trained LoRA tensors.
19
- - `mask_conditioning.safetensors` and `mask_conditioning_config.json`: trained explicit region conditioning.
20
- - `trainable_config.json` and `fingerprint.json`: original configuration and identity contracts.
21
- - `training_state.pt`: full optimizer, scheduler, sampler and rank RNG state for resuming the original E3 recipe.
22
- - `checkpoint_files.json`: exact file sizes and SHA256 digests.
23
-
24
- The checkpoint does not include the frozen base model or training data. Use the compatible released editMGT base model and the source repository's custom loader, supplying this downloaded folder as `trainable_state_path`. Preserve all checkpoint files together. The JSON configurations retain their original machine paths for reproducibility; data and asset paths need to be remapped on another machine, and strict resume identity checks must be honored.
25
-
26
- This is an exploratory epoch1 checkpoint. Teacher DEV128 inference has completed, but the comparative DiMO student/EMA quality evaluation is still incomplete; no formal quality acceptance is claimed. Corpus audits were skipped at the user's request, while frozen-manifest and runtime file integrity checks were retained.
27
-
28
- Download with the official Hugging Face CLI (authenticate to access this private repository):
29
-
30
- ```bash
31
- hf download TTangenty/editMGT-E3-epoch1-step6250 --local-dir /tmp/editmgt-e3-epoch1
32
- ```
 
5
  - lora
6
  - checkpoint
7
  ---
8
+ # editmgt-e3
9
 
10
+ Private checkpoint collection for the editMGT E3 and DiMO experiments. The frozen base model and datasets are not included. Each checkpoint directory contains its weights, the matching experiment metadata, and a README.
11
 
12
+ ## Checkpoints
13
 
14
+ - [`checkpoint-6250/`](./tree/main/checkpoint-6250): E3 epoch 1 teacher (200,000 sample exposures), including adapter/region weights and E3 optimizer/scheduler/sampler state. Source branch `fix/e3-quality-pipeline-v1`, commit `36dff5ff9ad10cee3d3910eb980932ec0a251b62`.
15
+ - [`checkpoint-7500/`](./tree/main/checkpoint-7500): DiMO student/auxiliary/EMA checkpoint (7,500 optimizer steps; 60,000 sample exposures), including strict-resume state. Initialized from E3 checkpoint-6250. Source branch `exp/dimo-e3-step3125-8gpu-v1`, commit `c4f92a7c696150c40f07959c72aa435cb6f67624`.
16
 
17
+ Both are intermediate research checkpoints. DiMO is still training; its checkpoint is not a completed epoch or a quality-validated release. Neither folder contains the frozen base model or training data. Checkpoint configurations preserve experiment provenance and bind resume to the original data/config/model identities; remap paths only under the source repository's documented identity rules.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
checkpoint-6250/README.md ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ # E3 checkpoint 6250
2
+
3
+ E3 epoch 1 teacher after 200,000 sample exposures. Includes adapter_model.safetensors, mask_conditioning.safetensors, configuration, fingerprint and complete E3 training_state.pt. Frozen backbone and data are not included. Source commit 36dff5ff9ad10cee3d3910eb980932ec0a251b62. Exploratory checkpoint; teacher DEV128 completed.
adapter_model.safetensors β†’ checkpoint-6250/adapter_model.safetensors RENAMED
File without changes
checkpoint_files.json β†’ checkpoint-6250/checkpoint_files.json RENAMED
File without changes
fingerprint.json β†’ checkpoint-6250/fingerprint.json RENAMED
File without changes
mask_conditioning.safetensors β†’ checkpoint-6250/mask_conditioning.safetensors RENAMED
File without changes
mask_conditioning_config.json β†’ checkpoint-6250/mask_conditioning_config.json RENAMED
File without changes
trainable_config.json β†’ checkpoint-6250/trainable_config.json RENAMED
File without changes
training_state.pt β†’ checkpoint-6250/training_state.pt RENAMED
File without changes
{dimo-e3-step6250-epoch1/checkpoint-7500 β†’ checkpoint-7500}/CHECKPOINT_READY.json RENAMED
File without changes
checkpoint-7500/README.md ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ # DiMO checkpoint 7500
2
+
3
+ DiMO student/auxiliary/EMA and full training state after 7,500 updates / 60,000 sample exposures, initialized from E3 epoch1 checkpoint-6250. Includes optimizer/scheduler, sampler cursor and distributed rank RNG for strict resume with the matching config, teacher, code and data identities. Frozen backbone and data are not included. This is an intermediate checkpoint while the full 25,000-step run continues; no final quality claim. Source commit c4f92a7c696150c40f07959c72aa435cb6f67624.
{dimo-e3-step6250-epoch1/checkpoint-7500 β†’ checkpoint-7500}/checkpoint_info.json RENAMED
File without changes
{dimo-e3-step6250-epoch1/checkpoint-7500 β†’ checkpoint-7500}/component_identity.json RENAMED
File without changes
{dimo-e3-step6250-epoch1/checkpoint-7500 β†’ checkpoint-7500}/dimo_config.yaml RENAMED
File without changes
{dimo-e3-step6250-epoch1/checkpoint-7500 β†’ checkpoint-7500}/dimo_trainable_parameter_report.json RENAMED
File without changes
{dimo-e3-step6250-epoch1/checkpoint-7500 β†’ checkpoint-7500}/run_manifest.json RENAMED
File without changes
{dimo-e3-step6250-epoch1/checkpoint-7500 β†’ checkpoint-7500}/training_state.pt RENAMED
File without changes
dimo-e3-step6250-epoch1/checkpoint-7500/README.md DELETED
@@ -1,7 +0,0 @@
1
- # DiMO E3 checkpoint at step 7500
2
-
3
- Training checkpoint saved after 7,500 optimizer updates and 60,000 sample exposures (0.3 epoch of the frozen 200,000-sample corpus). Initialized from E3 teacher checkpoint 6250 at `TTangenty/editMGT-E3-epoch1-step6250`.
4
-
5
- `training_state.pt` is the complete distributed checkpoint containing student and auxiliary trainable state, student EMA, optimizers, schedulers, sampler cursor, and rank RNG. The state is private to this experiment and bound to its exact teacher, Git revision, configuration, data manifest and 8-GPU world size. It is suitable for strict continuation under the matching experiment setup. It does not include the frozen base model or corpus.
6
-
7
- This is an intermediate training checkpoint, not a quality-validated release. The local training state was SHA256-checked before upload.