Add df3-stage1-r19 step 120000 (EMA shadow) safetensors + model card

README.md

@@ -0,0 +1,57 @@
+---
+tags:
+  - protein-structure-prediction
+  - alphafold3
+  - deepfold3
+---
+
+# DeepFold3
+
+Weights for DeepFold3, a PyTorch biomolecular complex structure predictor architecturally equivalent to AlphaFold 3.
+Each `.safetensors` file embeds its architectural `ModelConfig` as JSON in the header (`model_cfg_json`), so
+`deepfold3 inference` rebuilds the trained architecture without a sidecar config.
+
+## Checkpoints
+
+| File | Run | Step | Snapshot | Params | SHA-256 |
+| --- | --- | --- | --- | --- | --- |
+| `df3-stage1-r19/checkpoint_120000.safetensors` | df3-stage1-r19 | 120,000 | EMA shadow (decay 0.999) | 310,190,170 (fp32) | `aef0468336c0a0ed7ecdadbb5397cfbfe3de0c552b72957dc77b6d1fea29ba2e` |
+
+## df3-stage1-r19 @ 120k
+
+**Architecture (r19).** The trunk has no triangle attention and no single track: each PairFormer block is
+TriMulOutgoing + TriMulIncoming (plain AF3 TriangleMul) + FFN, following the pair-track design of
+arXiv:2510.18870. Diffusion conditioning does not consume the trunk single (`conditioning_use_trunk_single: false`).
+The confidence-head and template PairFormers are unchanged from AF3.
+
+| | |
+| --- | --- |
+| pair / single / MSA channels | 256 / 384 / 128 |
+| PairFormer layers | 48 |
+| MSA module layers | 4 |
+| Diffusion transformer blocks | 24 |
+| Confidence PairFormer layers | 4 |
+| Recycles | 10 |
+
+**Training.**
+
+- Stage 0 from scratch (`configs/df3-stage0-r19.yaml`) to step 71,000, then stage 1 (`configs/stage1.yaml`:
+  512-token crop, bond loss on, shape-complementarity loss 0.03) resumed from that checkpoint.
+- AdamW, lr 9e-4, EMA decay 0.999, 112 GPUs.
+- Dataset mix (weights at step 120k): OpenFold3 `pdb_training_set` 0.5, OF3 long monomers 0.2, AFDB (NVIDIA)
+  0.15, TEDdymer 0.1, OF3 short monomers 0.05. The weighted PDB set was `patchr` until step 20,000.
+- Step 120,000 is the optimizer step (`effective_step`). The checkpoint's `global_step` is 169,000 because stage 1
+  accumulates 2 micro-batches per step.
+
+**Export.** `deepfold3 export --snapshot auto` from `runs/df3-stage1-r19/checkpoint_120000.pt` (commit `131baf6`).
+The run is AdamW + EMA, so `auto` selects the EMA shadow (the deployment snapshot used by inference and eval).
+Verified bitwise-identical to the checkpoint's shadow overlay; loads into `DeepFold3` with no missing or
+unexpected keys.
+
+## Usage
+
+```bash
+hf download vv137/deepfold3 df3-stage1-r19/checkpoint_120000.safetensors --local-dir weights/
+deepfold3 inference --json input.json --output out/ \
+    --params weights/df3-stage1-r19/checkpoint_120000.safetensors
+```

df3-stage1-r19/checkpoint_120000.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:aef0468336c0a0ed7ecdadbb5397cfbfe3de0c552b72957dc77b6d1fea29ba2e
+size 1241052144