Update Protenix-RNA model card and validation figures

.gitattributes

@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+figures/validation_lddt_curve.png filter=lfs diff=lfs merge=lfs -text

README.md

@@ -11,7 +11,7 @@ tags:
 datasets:
 - LiteFold/PDB
 model-index:
-- name: protenix-rna-finetune-ema-s12999
+- name: protenix-rna
   results:
   - task:
       type: structure-prediction
@@ -23,26 +23,27 @@ model-index:
     - type: lddt_complex_best
       name: lDDT complex best
       value: 0.758663
+    - type: lddt_complex_mean
+      name: lDDT complex mean
+      value: 0.746286
     - type: lddt_complex_rank1
       name: lDDT complex rank1
       value: 0.746743
-    - type: validation_loss
-      name: Validation loss
-      value: 411.541803
 ---
 
-# Protenix RNA Fine-Tune EMA S12999
+# Protenix-RNA
 
-This repository contains a Protenix RNA fine-tuned PyTorch checkpoint selected from the EMA validation metric at training step 12,999. It is intended for use with the Protenix codebase rather than `transformers.AutoModel`.
+Protenix-RNA is a Protenix fine-tuned PyTorch checkpoint optimized for RNA structure prediction. It was selected by the EMA validation lDDT-complex best metric at training step 12,999 and is distributed as a native Protenix checkpoint for the Protenix codebase, not as a `transformers.AutoModel` package.
 
 ## Files
 
 | File | Description |
 |---|---|
-| `checkpoints/best_ema_0.999.pt` | Uploaded checkpoint from `output/protenix_rna_resume_opt_b32_lr5e5_s9500_to_s20000_20260522_231945/checkpoints/best_ema_0.999.pt`. |
-| `config.yaml` | Resolved config for the fine-tuning/evaluation run. |
-| `validation_comparison.csv` | Base vs fine-tuned validation metrics. |
-| `checkpoint_info.json` | Local source path, checkpoint step, and artifact metadata. |
+| `checkpoints/best_ema_0.999.pt` | EMA checkpoint selected at step 12,999. |
+| `config.yaml` | Resolved fine-tuning/evaluation config. |
+| `validation_comparison.csv` | lDDT-only validation comparison against the base and previous fine-tuned checkpoints. |
+| `checkpoint_info.json` | Source path, checkpoint step, and artifact metadata. |
+| `figures/` | Validation comparison and lDDT progression plots. |
 
 The checkpoint is a `torch.load(..., weights_only=False)` dictionary with keys `model`, `optimizer`, `scheduler`, and `step`. The stored step is `12999`.
 
@@ -58,38 +59,38 @@ The checkpoint is a `torch.load(..., weights_only=False)` dictionary with keys `
 - EMA decay: 0.999
 - Selection metric: `rna_finetune_val/ema0.999_lddt/complex/best.avg`, maximize
 
-## Validation Comparison
+## Validation
 
-Higher is better for lDDT metrics. Lower is better for loss metrics.
+Higher is better for all metrics shown here.
 
-| Metric | Base | Prior fine-tune s9499 | Uploaded EMA s12999 | Delta vs base | Delta vs s9499 |
+| Metric | Base Protenix | Prior FT s9499 | Protenix-RNA s12999 | Gain vs base | Gain vs s9499 |
 |---|---:|---:|---:|---:|---:|
-| loss | 1249.9014 | 890.1429 | 411.5418 | -838.3596 | -478.6011 |
-| weighted_mse | 1247.7881 | 888.5646 | 410.0376 | -837.7506 | -478.5270 |
-| mse | 311.9470 | 222.1411 | 102.5094 | -209.4376 | -119.6318 |
-| smooth_lddt_loss | 0.5282 | 0.3945 | 0.3759 | -0.1522 | -0.0186 |
-| lddt_best | 0.5558 | 0.7395 | 0.7587 | +0.2029 | +0.0192 |
-| lddt_mean | 0.5420 | 0.7261 | 0.7463 | +0.2043 | +0.0202 |
-| lddt_rank1 | 0.5417 | 0.7254 | 0.7467 | +0.2050 | +0.0214 |
-| pde | 2.8927 | 1.9580 | 2.1069 | -0.7858 | +0.1489 |
-| pae | 2.8719 | 3.6604 | 3.8774 | +1.0055 | +0.2170 |
-
-Validation settings: RNA validation split, seed 42, bf16, `N_sample=5`, `N_step=20`, `N_cycle=4`, `max_n_token=768`, RNA MSA enabled. The uploaded step 12,999 values come from the EMA validation loop that produced the best checkpoint. The base and prior fine-tune rows come from the standalone full-validation runs in this workspace.
+| lDDT best | 0.5558 | 0.7395 | 0.7587 | +0.2029 | +0.0192 |
+| lDDT mean | 0.5420 | 0.7261 | 0.7463 | +0.2043 | +0.0202 |
+| lDDT rank1 | 0.5417 | 0.7254 | 0.7467 | +0.2050 | +0.0214 |
+
+Validation settings: RNA validation split, seed 42, bf16, `N_sample=5`, `N_step=20`, `N_cycle=4`, `max_n_token=768`, RNA MSA enabled. The step 12,999 values come from the EMA validation loop that produced the uploaded checkpoint.
+
+![RNA validation lDDT comparison](figures/lddt_comparison.png)
+
+![Uploaded checkpoint lDDT gain](figures/lddt_gain.png)
+
+![RNA validation lDDT during fine-tuning](figures/validation_lddt_curve.png)
 
 ## Usage
 
 Download the checkpoint and point Protenix at it with `--load_params_only true`:
 
 ```bash
-hf download LiteFold/protenix-rna-finetune-ema-s12999 \
+hf download LiteFold/protenix-rna \
   checkpoints/best_ema_0.999.pt \
-  --local-dir ./protenix-rna-finetune-ema-s12999
+  --local-dir ./protenix-rna
 ```
 
 Example evaluation invocation inside the Protenix checkout:
 
 ```bash
-LOAD_CHECKPOINT_PATH=./protenix-rna-finetune-ema-s12999/checkpoints/best_ema_0.999.pt \
+LOAD_CHECKPOINT_PATH=./protenix-rna/checkpoints/best_ema_0.999.pt \
 VAL_MAX_N_TOKEN=768 \
 VAL_LIMIT=-1 \
 N_SAMPLE=5 \
@@ -110,4 +111,4 @@ step = ckpt["step"]
 
 ## Limitations
 
-This is a research checkpoint specialized for the local RNA fine-tuning setup. It has not been packaged as a standalone Transformers model and should be evaluated with the same Protenix code/configuration family used for training.
+This is a research checkpoint specialized for the RNA fine-tuning setup above. It has not been converted into a standalone Transformers model and should be evaluated with the same Protenix code/configuration family used for training.

checkpoint_info.json

@@ -1,5 +1,6 @@
 {
   "checkpoint_name": "best_ema_0.999.pt",
+  "repo_id": "LiteFold/protenix-rna",
   "source_path": "output/protenix_rna_resume_opt_b32_lr5e5_s9500_to_s20000_20260522_231945/checkpoints/best_ema_0.999.pt",
   "path_in_repo": "checkpoints/best_ema_0.999.pt",
   "size_bytes": 4427468333,

validation_comparison.csv

@@ -1,10 +1,4 @@
-metric,base_default_v1,prior_finetune_ema_s9499,uploaded_ema_s12999,delta_s12999_vs_base,delta_s12999_vs_s9499,higher_is_better
-loss,1249.901394,890.142949,411.541803,-838.359591,-478.601146,false
-weighted_mse,1247.788115,888.564566,410.037558,-837.750556,-478.527007,false
-mse,311.947029,222.141141,102.509390,-209.437639,-119.631752,false
-smooth_lddt_loss,0.528177,0.394495,0.375942,-0.152235,-0.018554,false
+metric,base_default_v1,prior_finetune_ema_s9499,protenix_rna_ema_s12999,gain_s12999_vs_base,gain_s12999_vs_s9499,higher_is_better
 lddt_best,0.555753,0.739509,0.758663,0.202910,0.019154,true
 lddt_mean,0.541968,0.726095,0.746286,0.204318,0.020192,true
 lddt_rank1,0.541723,0.725381,0.746743,0.205021,0.021363,true
-pde,2.892729,1.958018,2.106942,-0.785787,0.148924,false
-pae,2.871937,3.660412,3.877398,1.005460,0.216986,false