Group checkpoints by experiment family (qm9/, geom/, dna/), name folders by role, drop the diagonal-only QM9 student

README.md

@@ -28,13 +28,24 @@ python scripts/download_artifacts.py --only geom    # or one family: qm9, dna, g
 
 ## Contents
 
+One folder of checkpoints per experiment family, with the same role names in each:
+
 | path | what | size |
 |---|---|---|
-| `checkpoints/base_flow/` | QM9 base flow model (TABASCO discrete-flow-matching transformer) | 4 MB |
-| `checkpoints/mfm_student/`, `checkpoints/mfm_student_ce_esd/` | QM9 multiMFM students (the CE/ESD student is the one behind the QM9 table) | 2 MB |
-| `checkpoints/regressors/` | TFG-Flow QM9 property regressors: guide (steering) and oracle (scoring) | 41 MB |
-| `checkpoints/geom/` | GEOM-Drugs base model, multiMFM student and FP2 fingerprint guide head | 422 MB |
-| `checkpoints/dna/` | DNA base DFM denoisers and dMFM students (L = 50, 100, 200, 400), C0 guide and oracle | 499 MB |
+| **`checkpoints/qm9/`** | **QM9 property steering** | **47 MB** |
+| `  base_model/` | base flow model (TABASCO discrete-flow-matching transformer) | 4 MB |
+| `  mfm_student_ce_esd/` | multiMFM student: cross-entropy atom loss + ESD consistency; sampled with 4 flow-map jumps | 1.5 MB |
+| `  property_regressors/` | TFG-Flow property regressors: guide (steering) and oracle (scoring only) | 41 MB |
+| **`checkpoints/geom/`** | **GEOM-Drugs structural-fingerprint steering** | **422 MB** |
+| `  base_model/` | base flow model (Lightning checkpoint, EMA weights) and its Hydra config | 297 MB |
+| `  mfm_student/` | multiMFM student | 121 MB |
+| `  fingerprint_head/` | FP2 fingerprint guide head (TFG-Flow EGNN architecture), the steering reward | 4 MB |
+| **`checkpoints/dna/`** | **DNA cyclizability steering**, sequence lengths L = 50, 100, 200, 400 | **499 MB** |
+| `  base_model/L*/` | base discrete-flow-matching denoisers | 140 MB |
+| `  dmfm_student/L*/` | dMFM students (diagonal + ESD), used as one-step posterior maps | 159 MB |
+| `  dmfm_student_4step/L*/` | dMFM students specialised for 4-step composition | 159 MB |
+| `  dmfm_student_L400_parent/` | the L400 student before its diagonal-only finetune | 40 MB |
+| `  c0_regressors/{guide,oracle}/` | cyclizability (C0) regressors, trained on disjoint parent halves | 1 MB |
 | `data/dna/` | loop-seq yeast cyclizability windows with parent-disjoint splits | 385 MB |
 | `data/geom/geom_fp/` | FP2 fingerprints of the GEOM-Drugs val split (fingerprint-steering targets) | 173 MB |
 

checkpoints/mfm_student/student_step_1000.pt

@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:67953452e2a19f3d80d579ddce96ea27dce237a58c2fad9e2b4662ef41bd3cd0
-size 1488848