Add Gyuyig vs Tsugdri binary classifier weights, metrics, and training history

README.md

@@ -0,0 +1,149 @@
+---
+language:
+  - bo
+license: apache-2.0
+tags:
+  - image-classification
+  - tibetan
+  - script-classification
+  - dinov3
+  - binary
+library_name: transformers
+pipeline_tag: image-classification
+base_model: facebook/dinov3-vits16-pretrain-lvd1689m
+datasets:
+  - BDRC/gyuyig-tsugdri-binary-balanced-script-classification-dataset
+metrics:
+  - f1
+  - accuracy
+  - auc
+---
+
+# Gyuyig vs Tsugdri Binary Script Classifier (DINOv3 ViT-S)
+
+Fine-tuned [DINOv3 ViT-S](https://huggingface.co/facebook/dinov3-vits16-pretrain-lvd1689m) for parent script classification:
+
+**Gyuyig**, **Tsugdri**
+
+**Experiment:** `dinov3_gyuyig_tsugdri_sub_warmstart` (`gyuyig_tsugdri_binary_classification`)  
+**Pooling:** ViT **CLS token** (`last_hidden_state[:, 0, :]`)  
+**Weights:** `final_model.pt` (best validation macro-F1 across stages A/B/C)
+
+**Warm-start:** [BDRC/4-class-balanced-script-classifier](https://huggingface.co/BDRC/4-class-balanced-script-classifier) (`final_model.pt` — prior test acc 82.6%, macro-F1 0.833)
+
+## Data
+
+| Split | Source |
+|-------|--------|
+| Train / val / test | [BDRC/gyuyig-tsugdri-binary-balanced-script-classification-dataset](https://huggingface.co/datasets/BDRC/gyuyig-tsugdri-binary-balanced-script-classification-dataset) |
+
+Test split: balanced benchmark (60 images per parent class, held out of training).
+
+## Preprocessing
+
+| Split | Mode | Size |
+|-------|------|-----:|
+| train | `resize_letterbox` | 448 |
+| val | `resize_letterbox` | 448 |
+| test | `resize_letterbox` | 448 |
+
+## Validation metrics (n=60)
+
+| Metric | Value |
+|--------|------:|
+| Accuracy | 85.0% |
+| Macro F1 | 0.850 |
+| Weighted F1 | 0.850 |
+| AUC-ROC | 0.902 |
+| Loss | 0.4520 |
+
+**Best checkpoint:** `best_stage_c_last_blocks.pt` epoch 7 val macro-F1 0.850
+
+### Per-class (validation)
+
+```
+precision    recall  f1-score   support
+
+      Gyuyig       0.89      0.80      0.84        30
+     Tsugdri       0.82      0.90      0.86        30
+
+    accuracy                           0.85        60
+   macro avg       0.85      0.85      0.85        60
+weighted avg       0.85      0.85      0.85        60
+```
+
+## Test / benchmark metrics (n=120)
+
+| Metric | Value |
+|--------|------:|
+| Accuracy | 80.8% |
+| Macro F1 | 0.808 |
+| Weighted F1 | 0.808 |
+| AUC-ROC | 0.868 |
+| Loss | 0.5354 |
+
+### Per-class (test)
+
+```
+precision    recall  f1-score   support
+
+      Gyuyig       0.79      0.83      0.81        60
+     Tsugdri       0.82      0.78      0.80        60
+
+    accuracy                           0.81       120
+   macro avg       0.81      0.81      0.81       120
+weighted avg       0.81      0.81      0.81       120
+```
+
+## Training
+
+| Stage | Epochs | LR head | LR backbone | Unfrozen blocks |
+|-------|-------:|--------:|------------:|----------------:|
+| A | 7 | 0.0005 | — | 0 |
+| B | 10 | 0.0001 | 1e-05 | 4 |
+| C | 12 | 5e-05 | 1.5e-05 | 8 |
+
+| Setting | Value |
+|---------|-------|
+| Scheduler | `cosine_warmup` |
+| Class weights | `custom` |
+| Label smoothing | 0.05 |
+| Dropout | 0.1 |
+
+## Confusion matrix (test)
+
+![Confusion matrix](confusion_matrix.png)
+
+| True \ Pred | Gyuyig | Tsugdri |
+|---|---:|---:|
+| **Gyuyig** | 50 | 10 |
+| **Tsugdri** | 13 | 47 |
+
+## Files
+
+| File | Description |
+|------|-------------|
+| `final_model.pt` | Best val-F1 weights + label maps |
+| `results.json` | Full metrics, history, warm-start info |
+| `config.yaml` | Training config |
+| `model_card.json` | Summary metadata |
+| `confusion_matrix.json` / `.png` | Test CM |
+| `training_history.png` | Stage loss / val F1 curves |
+| `split_stats.json` / `.md` | Per-class split counts |
+| `inference.py` | Classify image paths |
+| `requirements-inference.txt` | Pip deps |
+
+## Inference
+
+```bash
+pip install -r requirements-inference.txt
+python inference.py --checkpoint final_model.pt --image path/to/page.jpg --preprocess resize_letterbox --preprocess-size 448
+```
+
+## Reproduce training
+
+```bash
+python experiments/gyuyig-tsugdri/train.py --config experiments/gyuyig-tsugdri/config_warmstart.yaml
+```
+
+**Model repo:** [BDRC/gyuyig-tsugdri-binary-script-classifier](https://huggingface.co/BDRC/gyuyig-tsugdri-binary-script-classifier)

config.yaml

@@ -0,0 +1,53 @@
+experiment: dinov3_gyuyig_tsugdri_sub_warmstart
+task: gyuyig_tsugdri_binary_classification
+
+balanced_dataset_repo: BDRC/gyuyig-tsugdri-binary-balanced-script-classification-dataset
+val_ratio: 0.15
+
+warmstart_repo: BDRC/4-class-balanced-script-classifier
+warmstart_checkpoint_file: final_model.pt
+
+output_dir: checkpoints
+model_id: facebook/dinov3-vits16-pretrain-lvd1689m
+seed: 42
+batch_size: 16
+num_workers: 8
+no_amp: false
+no_weighted_sampler: true
+skip_stage_c: false
+gradient_checkpointing: true
+
+train_preprocess: resize_letterbox
+val_preprocess: resize_letterbox
+test_preprocess: resize_letterbox
+preprocess_size: 448
+
+pooling: cls_token
+
+epochs_a: 7
+epochs_b: 10
+epochs_c: 12
+unfreeze_blocks_b: 4
+unfreeze_blocks_c: 8
+
+lr_head_a: 5.0e-4
+lr_backbone_b: 1.0e-5
+lr_head_b: 1.0e-4
+lr_backbone_c: 1.5e-5
+lr_head_c: 5.0e-5
+
+scheduler: cosine_warmup
+warmup_epochs_a: 0
+warmup_epochs_b: 1
+warmup_epochs_c: 1
+warmup_start_factor: 0.01
+min_lr_ratio: 0.01
+
+weight_decay: 0.02
+grad_clip_norm: 1.0
+dropout: 0.1
+label_smoothing: 0.05
+class_weight_mode: sqrt_inverse_freq
+
+early_stop_patience: 6
+early_stop_min_delta: 0.001

confusion_matrix.json

@@ -0,0 +1,41 @@
+{
+  "split": "test",
+  "labels": [
+    "Gyuyig",
+    "Tsugdri"
+  ],
+  "matrix": [
+    [
+      50,
+      10
+    ],
+    [
+      13,
+      47
+    ]
+  ],
+  "test_metrics": {
+    "loss": 0.5353951374689738,
+    "accuracy": 0.8083333333333333,
+    "macro_f1": 0.8082134667500521,
+    "weighted_f1": 0.8082134667500521,
+    "auc_roc": 0.8677777777777778
+  },
+  "val_metrics": {
+    "loss": 0.45204214652379354,
+    "accuracy": 0.85,
+    "macro_f1": 0.849624060150376,
+    "weighted_f1": 0.849624060150376,
+    "auc_roc": 0.9022222222222223
+  },
+  "preprocess": {
+    "train": "resize_letterbox",
+    "val": "resize_letterbox",
+    "test": "resize_letterbox",
+    "size": 448
+  },
+  "train_dataset": "BDRC/gyuyig-tsugdri-binary-balanced-script-classification-dataset",
+  "benchmark_per_parent": 60,
+  "experiment": "dinov3_gyuyig_tsugdri_sub_warmstart",
+  "repo_id": "BDRC/gyuyig-tsugdri-binary-script-classifier"
+}

final_model.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6753c4b49f6fa52b6d4580b926070cdea8dba7908f0e3a562f60dd42512e3148
+size 86670182

inference.py

@@ -0,0 +1,132 @@
+#!/usr/bin/env python3
+"""Standalone DINOv3 script classifier inference (copied to Hub model repo as ``inference.py``)."""
+
+from __future__ import annotations
+
+import argparse
+from pathlib import Path
+
+import torch
+import torch.nn as nn
+from PIL import Image
+from transformers import AutoImageProcessor, AutoModel
+
+DINOV3_MODEL_ID = "facebook/dinov3-vits16-pretrain-lvd1689m"
+
+
+class DINOv3Classifier(nn.Module):
+    def __init__(self, model_id: str, num_classes: int, dropout: float = 0.1):
+        super().__init__()
+        self.backbone = AutoModel.from_pretrained(model_id)
+        hidden = self.backbone.config.hidden_size
+        self.head = nn.Sequential(
+            nn.LayerNorm(hidden),
+            nn.Dropout(dropout),
+            nn.Linear(hidden, 128),
+            nn.GELU(),
+            nn.Dropout(dropout),
+            nn.Linear(128, num_classes),
+        )
+
+    def forward(self, pixel_values):
+        out = self.backbone(pixel_values=pixel_values)
+        cls = out.last_hidden_state[:, 0, :]
+        return self.head(cls)
+
+
+def _resize_short_edge(img: Image.Image, target: int) -> Image.Image:
+    w, h = img.size
+    if h <= w:
+        new_h = target
+        new_w = max(1, int(w * target / h))
+    else:
+        new_w = target
+        new_h = max(1, int(h * target / w))
+    return img.resize((new_w, new_h), Image.BICUBIC)
+
+
+def _center_crop(img: Image.Image, size: int = 224) -> Image.Image:
+    img = _resize_short_edge(img, size)
+    w, h = img.size
+    left = max(0, (w - size) // 2)
+    top = max(0, (h - size) // 2)
+    crop = img.crop((left, top, left + size, top + size))
+    if crop.size != (size, size):
+        padded = Image.new("RGB", (size, size), (255, 255, 255))
+        padded.paste(crop, (0, 0))
+        return padded
+    return crop
+
+
+def apply_preprocess(img: Image.Image, mode: str | None, *, size: int = 224) -> Image.Image:
+    if not mode or mode == "none":
+        return img
+    if mode in ("center_crop", "center_crop_whole_page"):
+        return _center_crop(img, size)
+    raise ValueError(f"Unknown preprocess mode: {mode!r}")
+
+
+def label_order(ckpt: dict) -> list[str]:
+    idx = ckpt.get("idx_to_label") or {}
+    if idx:
+        return [str(idx[k]) for k in sorted(idx.keys(), key=lambda x: int(x))]
+    raw = ckpt.get("label_to_idx") or {}
+    if raw:
+        return sorted(raw.keys(), key=lambda k: raw[k])
+    raise KeyError("checkpoint missing idx_to_label / label_to_idx")
+
+
+@torch.no_grad()
+def predict(model, processor, image_path: Path, device, *, preprocess: str | None, size: int):
+    img = Image.open(image_path).convert("RGB")
+    img = apply_preprocess(img, preprocess, size=size)
+    pv = processor(images=img, return_tensors="pt")["pixel_values"].to(device)
+    logits = model(pv)
+    probs = torch.softmax(logits, dim=1).squeeze(0).cpu()
+    pred = int(probs.argmax())
+    return pred, probs.tolist()
+
+
+def main() -> None:
+    ap = argparse.ArgumentParser(description="DINOv3 Tibetan script page classifier")
+    ap.add_argument(
+        "--checkpoint",
+        type=Path,
+        default=Path("final_model.pt"),
+        help="Weights file (default: final_model.pt in cwd)",
+    )
+    ap.add_argument("--image", type=Path, nargs="+", required=True)
+    ap.add_argument(
+        "--preprocess",
+        default="none",
+        help="none (full page, DINO processor resize) | center_crop (224 short edge)",
+    )
+    ap.add_argument("--preprocess-size", type=int, default=224)
+    ap.add_argument("--model-id", default=DINOV3_MODEL_ID)
+    args = ap.parse_args()
+
+    ckpt = torch.load(args.checkpoint, map_location="cpu", weights_only=False)
+    classes = label_order(ckpt)
+    idx_to_label = {i: lab for i, lab in enumerate(classes)}
+
+    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+    model = DINOv3Classifier(args.model_id, num_classes=len(classes)).to(device)
+    model.load_state_dict(ckpt["model_state_dict"])
+    model.eval()
+    processor = AutoImageProcessor.from_pretrained(args.model_id)
+    prep = None if args.preprocess in ("none", "") else args.preprocess
+
+    for path in args.image:
+        pred, probs = predict(
+            model, processor, path, device, preprocess=prep, size=args.preprocess_size
+        )
+        name = idx_to_label[pred]
+        conf = probs[pred]
+        print(f"{path.name}: {name} ({conf:.3f})")
+        top3 = sorted(enumerate(probs), key=lambda x: -x[1])[:3]
+        for i, p in top3:
+            print(f"  {idx_to_label[i]:16s} {p:.3f}")
+
+
+if __name__ == "__main__":
+    main()

model_card.json

@@ -0,0 +1,86 @@
+{
+  "repo_id": "BDRC/gyuyig-tsugdri-binary-script-classifier",
+  "train_dataset_id": "BDRC/gyuyig-tsugdri-binary-balanced-script-classification-dataset",
+  "warmstart_repo": "BDRC/4-class-balanced-script-classifier",
+  "task": "gyuyig_tsugdri_binary_classification",
+  "experiment": "dinov3_gyuyig_tsugdri_sub_warmstart",
+  "classes": [
+    "Gyuyig",
+    "Tsugdri"
+  ],
+  "pooling": "cls_token",
+  "preprocess": {
+    "train": "resize_letterbox",
+    "val": "resize_letterbox",
+    "test": "resize_letterbox",
+    "size": 448
+  },
+  "warmstart": {
+    "warmstart_repo": "BDRC/4-class-balanced-script-classifier",
+    "warmstart_checkpoint": null,
+    "warmstart_checkpoint_file": "final_model.pt",
+    "checkpoint_test_metrics": {
+      "loss": 0.6574946736847913,
+      "accuracy": 0.825925925925926,
+      "macro_f1": 0.8326187473728457,
+      "weighted_f1": 0.82908384875598
+    },
+    "warmstart_pooling": "cls_token"
+  },
+  "training_config": {
+    "lr_head_a": 0.0005,
+    "lr_backbone_b": 1e-05,
+    "lr_head_b": 0.0001,
+    "lr_backbone_c": 1.5e-05,
+    "lr_head_c": 5e-05,
+    "epochs_a": 7,
+    "epochs_b": 10,
+    "epochs_c": 12,
+    "unfreeze_blocks_b": 4,
+    "unfreeze_blocks_c": 8,
+    "skip_stage_c": false,
+    "scheduler": "cosine_warmup",
+    "warmup_epochs_a": 0,
+    "warmup_epochs_b": 1,
+    "warmup_epochs_c": 1,
+    "warmup_start_factor": 0.01,
+    "min_lr_ratio": 0.01,
+    "weight_decay": 0.02,
+    "grad_clip_norm": 1.0,
+    "class_weight_mode": "custom",
+    "class_weights": [
+      1.0,
+      1.0
+    ],
+    "label_smoothing": 0.05,
+    "dropout": 0.1,
+    "early_stop_patience": 6,
+    "early_stop_min_delta": 0.001,
+    "resolved_class_weights": [
+      1.0,
+      1.0
+    ],
+    "pooling": "cls_token",
+    "lambda_consistency": 0.0,
+    "flip_op": "rot180"
+  },
+  "best_checkpoint": {
+    "path": "best_stage_c_last_blocks.pt",
+    "epoch": 7,
+    "val_macro_f1": 0.849624060150376
+  },
+  "val_metrics": {
+    "loss": 0.45204214652379354,
+    "accuracy": 0.85,
+    "macro_f1": 0.849624060150376,
+    "weighted_f1": 0.849624060150376,
+    "auc_roc": 0.9022222222222223
+  },
+  "test_metrics": {
+    "loss": 0.5353951374689738,
+    "accuracy": 0.8083333333333333,
+    "macro_f1": 0.8082134667500521,
+    "weighted_f1": 0.8082134667500521,
+    "auc_roc": 0.8677777777777778
+  }
+}

requirements-inference.txt

@@ -0,0 +1,3 @@
+torch>=2.0
+transformers>=4.40
+pillow>=10.0

results.json

@@ -0,0 +1,454 @@
+{
+  "experiment": "dinov3_gyuyig_tsugdri_sub_warmstart",
+  "run_subdir": null,
+  "task": "gyuyig_tsugdri_binary_classification",
+  "balanced_parquet_dir": null,
+  "balanced_dataset_repo": "BDRC/gyuyig-tsugdri-binary-balanced-script-classification-dataset",
+  "data_source": "hf",
+  "benchmark_dataset_repo": null,
+  "benchmark_per_parent": 60,
+  "preprocess": {
+    "train": "resize_letterbox",
+    "val": "resize_letterbox",
+    "test": "resize_letterbox",
+    "size": 448
+  },
+  "pooling": "cls_token",
+  "training_config": {
+    "lr_head_a": 0.0005,
+    "lr_backbone_b": 1e-05,
+    "lr_head_b": 0.0001,
+    "lr_backbone_c": 1.5e-05,
+    "lr_head_c": 5e-05,
+    "epochs_a": 7,
+    "epochs_b": 10,
+    "epochs_c": 12,
+    "unfreeze_blocks_b": 4,
+    "unfreeze_blocks_c": 8,
+    "skip_stage_c": false,
+    "scheduler": "cosine_warmup",
+    "warmup_epochs_a": 0,
+    "warmup_epochs_b": 1,
+    "warmup_epochs_c": 1,
+    "warmup_start_factor": 0.01,
+    "min_lr_ratio": 0.01,
+    "weight_decay": 0.02,
+    "grad_clip_norm": 1.0,
+    "class_weight_mode": "custom",
+    "class_weights": [
+      1.0,
+      1.0
+    ],
+    "label_smoothing": 0.05,
+    "dropout": 0.1,
+    "early_stop_patience": 6,
+    "early_stop_min_delta": 0.001,
+    "resolved_class_weights": [
+      1.0,
+      1.0
+    ],
+    "pooling": "cls_token",
+    "lambda_consistency": 0.0,
+    "flip_op": "rot180"
+  },
+  "resolved_class_weights": [
+    1.0,
+    1.0
+  ],
+  "warmstart": {
+    "warmstart_repo": "BDRC/4-class-balanced-script-classifier",
+    "warmstart_checkpoint": null,
+    "warmstart_checkpoint_file": "final_model.pt",
+    "checkpoint_test_metrics": {
+      "loss": 0.6574946736847913,
+      "accuracy": 0.825925925925926,
+      "macro_f1": 0.8326187473728457,
+      "weighted_f1": 0.82908384875598
+    },
+    "warmstart_pooling": "cls_token"
+  },
+  "checkpoint_selection": {
+    "method": "best_val_macro_f1"
+  },
+  "best_checkpoint": {
+    "path": "best_stage_c_last_blocks.pt",
+    "epoch": 7,
+    "val_macro_f1": 0.849624060150376
+  },
+  "val_metrics": {
+    "loss": 0.45204214652379354,
+    "accuracy": 0.85,
+    "macro_f1": 0.849624060150376,
+    "weighted_f1": 0.849624060150376,
+    "auc_roc": 0.9022222222222223
+  },
+  "val_confusion_matrix": [
+    [
+      24,
+      6
+    ],
+    [
+      3,
+      27
+    ]
+  ],
+  "val_report": "              precision    recall  f1-score   support\n\n      Gyuyig       0.89      0.80      0.84        30\n     Tsugdri       0.82      0.90      0.86        30\n\n    accuracy                           0.85        60\n   macro avg       0.85      0.85      0.85        60\nweighted avg       0.85      0.85      0.85        60\n",
+  "test_metrics": {
+    "loss": 0.5353951374689738,
+    "accuracy": 0.8083333333333333,
+    "macro_f1": 0.8082134667500521,
+    "weighted_f1": 0.8082134667500521,
+    "auc_roc": 0.8677777777777778
+  },
+  "test_confusion_matrix": [
+    [
+      50,
+      10
+    ],
+    [
+      13,
+      47
+    ]
+  ],
+  "test_report": "              precision    recall  f1-score   support\n\n      Gyuyig       0.79      0.83      0.81        60\n     Tsugdri       0.82      0.78      0.80        60\n\n    accuracy                           0.81       120\n   macro avg       0.81      0.81      0.81       120\nweighted avg       0.81      0.81      0.81       120\n",
+  "history": {
+    "stage_a": [
+      {
+        "epoch": 1,
+        "train_loss": 0.868421577271961,
+        "train_acc": 0.39880952380952384,
+        "lr_head": 5e-06,
+        "val_loss": 0.8907806555430094,
+        "val_accuracy": 0.3,
+        "val_macro_f1": 0.28,
+        "val_weighted_f1": 0.28
+      },
+      {
+        "epoch": 2,
+        "train_loss": 0.6925251398767743,
+        "train_acc": 0.5744047619047619,
+        "lr_head": 0.0005,
+        "val_loss": 0.6001647631327311,
+        "val_accuracy": 0.6166666666666667,
+        "val_macro_f1": 0.5911111111111111,
+        "val_weighted_f1": 0.5911111111111111
+      },
+      {
+        "epoch": 3,
+        "train_loss": 0.5732808538845607,
+        "train_acc": 0.6964285714285714,
+        "lr_head": 0.0004668412874366486,
+        "val_loss": 0.5495736837387085,
+        "val_accuracy": 0.7166666666666667,
+        "val_macro_f1": 0.7165879410947485,
+        "val_weighted_f1": 0.7165879410947485
+      },
+      {
+        "epoch": 4,
+        "train_loss": 0.5409448828016009,
+        "train_acc": 0.7351190476190477,
+        "lr_head": 0.00037624999999999996,
+        "val_loss": 0.5596829652786255,
+        "val_accuracy": 0.7166666666666667,
+        "val_macro_f1": 0.7127569698676429,
+        "val_weighted_f1": 0.7127569698676429
+      },
+      {
+        "epoch": 5,
+        "train_loss": 0.517706477925891,
+        "train_acc": 0.7380952380952381,
+        "lr_head": 0.0002525,
+        "val_loss": 0.570094374815623,
+        "val_accuracy": 0.6833333333333333,
+        "val_macro_f1": 0.6723196320781835,
+        "val_weighted_f1": 0.6723196320781833
+      },
+      {
+        "epoch": 6,
+        "train_loss": 0.5244638267017546,
+        "train_acc": 0.7529761904761905,
+        "lr_head": 0.00012875000000000007,
+        "val_loss": 0.5446342428525289,
+        "val_accuracy": 0.7166666666666667,
+        "val_macro_f1": 0.7159565580618212,
+        "val_weighted_f1": 0.7159565580618211
+      },
+      {
+        "epoch": 7,
+        "train_loss": 0.49209926100004286,
+        "train_acc": 0.7797619047619048,
+        "lr_head": 3.815871256335142e-05,
+        "val_loss": 0.5444834152857463,
+        "val_accuracy": 0.7166666666666667,
+        "val_macro_f1": 0.7159565580618212,
+        "val_weighted_f1": 0.7159565580618211
+      }
+    ],
+    "stage_b": [
+      {
+        "epoch": 1,
+        "train_loss": 0.5235937989893413,
+        "train_acc": 0.75,
+        "lr_head": 1.0000000000000002e-06,
+        "lr_backbone": 1.0000000000000001e-07,
+        "val_loss": 0.5491747617721557,
+        "val_accuracy": 0.7166666666666667,
+        "val_macro_f1": 0.7165879410947485,
+        "val_weighted_f1": 0.7165879410947485
+      },
+      {
+        "epoch": 2,
+        "train_loss": 0.5453660388787588,
+        "train_acc": 0.7529761904761905,
+        "lr_head": 0.0001,
+        "lr_backbone": 1e-05,
+        "val_loss": 0.5492973128954569,
+        "val_accuracy": 0.7333333333333333,
+        "val_macro_f1": 0.7306397306397306,
+        "val_weighted_f1": 0.7306397306397305
+      },
+      {
+        "epoch": 3,
+        "train_loss": 0.49581416731788996,
+        "train_acc": 0.7916666666666666,
+        "lr_head": 9.701478472890248e-05,
+        "lr_backbone": 9.701478472890248e-06,
+        "val_loss": 0.538078244527181,
+        "val_accuracy": 0.7166666666666667,
+        "val_macro_f1": 0.7165879410947485,
+        "val_weighted_f1": 0.7165879410947485
+      },
+      {
+        "epoch": 4,
+        "train_loss": 0.4958783601011549,
+        "train_acc": 0.7678571428571429,
+        "lr_head": 8.84191999343894e-05,
+        "lr_backbone": 8.841919993438941e-06,
+        "val_loss": 0.5364969690640767,
+        "val_accuracy": 0.7333333333333333,
+        "val_macro_f1": 0.7321428571428572,
+        "val_weighted_f1": 0.7321428571428572
+      },
+      {
+        "epoch": 5,
+        "train_loss": 0.4718715477557409,
+        "train_acc": 0.7946428571428571,
+        "lr_head": 7.525e-05,
+        "lr_backbone": 7.525e-06,
+        "val_loss": 0.5267467339833577,
+        "val_accuracy": 0.7333333333333333,
+        "val_macro_f1": 0.7321428571428572,
+        "val_weighted_f1": 0.7321428571428572
+      },
+      {
+        "epoch": 6,
+        "train_loss": 0.4493213253361838,
+        "train_acc": 0.8244047619047619,
+        "lr_head": 5.909558479451306e-05,
+        "lr_backbone": 5.909558479451306e-06,
+        "val_loss": 0.5186658461888631,
+        "val_accuracy": 0.7166666666666667,
+        "val_macro_f1": 0.7159565580618212,
+        "val_weighted_f1": 0.7159565580618211
+      },
+      {
+        "epoch": 7,
+        "train_loss": 0.43445002465021043,
+        "train_acc": 0.8184523809523809,
+        "lr_head": 4.190441520548695e-05,
+        "lr_backbone": 4.190441520548696e-06,
+        "val_loss": 0.5154884020487468,
+        "val_accuracy": 0.7166666666666667,
+        "val_macro_f1": 0.7159565580618212,
+        "val_weighted_f1": 0.7159565580618211
+      },
+      {
+        "epoch": 8,
+        "train_loss": 0.42955496197655085,
+        "train_acc": 0.8273809523809523,
+        "lr_head": 2.5750000000000013e-05,
+        "lr_backbone": 2.575000000000001e-06,
+        "val_loss": 0.5194342851638794,
+        "val_accuracy": 0.7666666666666667,
+        "val_macro_f1": 0.7643097643097643,
+        "val_weighted_f1": 0.7643097643097643
+      },
+      {
+        "epoch": 9,
+        "train_loss": 0.4368310116586231,
+        "train_acc": 0.8154761904761905,
+        "lr_head": 1.2580800065610596e-05,
+        "lr_backbone": 1.2580800065610596e-06,
+        "val_loss": 0.5142494519551595,
+        "val_accuracy": 0.7333333333333333,
+        "val_macro_f1": 0.7321428571428572,
+        "val_weighted_f1": 0.7321428571428572
+      },
+      {
+        "epoch": 10,
+        "train_loss": 0.4370355563504355,
+        "train_acc": 0.8125,
+        "lr_head": 3.985215271097539e-06,
+        "lr_backbone": 3.985215271097539e-07,
+        "val_loss": 0.5129753907521566,
+        "val_accuracy": 0.7333333333333333,
+        "val_macro_f1": 0.7321428571428572,
+        "val_weighted_f1": 0.7321428571428572
+      }
+    ],
+    "stage_c": [
+      {
+        "epoch": 1,
+        "train_loss": 0.4291627889587766,
+        "train_acc": 0.8273809523809523,
+        "lr_head": 5.000000000000001e-07,
+        "lr_backbone": 1.5000000000000002e-07,
+        "val_loss": 0.5183799107869466,
+        "val_accuracy": 0.7666666666666667,
+        "val_macro_f1": 0.7643097643097643,
+        "val_weighted_f1": 0.7643097643097643
+      },
+      {
+        "epoch": 2,
+        "train_loss": 0.42413380600157236,
+        "train_acc": 0.8154761904761905,
+        "lr_head": 5e-05,
+        "lr_backbone": 1.5e-05,
+        "val_loss": 0.5204356749852498,
+        "val_accuracy": 0.7166666666666667,
+        "val_macro_f1": 0.7101449275362319,
+        "val_weighted_f1": 0.7101449275362319
+      },
+      {
+        "epoch": 3,
+        "train_loss": 0.37433484338578726,
+        "train_acc": 0.875,
+        "lr_head": 4.899745109695881e-05,
+        "lr_backbone": 1.4699235329087644e-05,
+        "val_loss": 0.48167786995569867,
+        "val_accuracy": 0.7666666666666667,
+        "val_macro_f1": 0.7643097643097643,
+        "val_weighted_f1": 0.7643097643097643
+      },
+      {
+        "epoch": 4,
+        "train_loss": 0.335602682970819,
+        "train_acc": 0.8839285714285714,
+        "lr_head": 4.6071024937571735e-05,
+        "lr_backbone": 1.3821307481271522e-05,
+        "val_loss": 0.45847638448079425,
+        "val_accuracy": 0.8166666666666667,
+        "val_macro_f1": 0.8153846153846154,
+        "val_weighted_f1": 0.8153846153846154
+      },
+      {
+        "epoch": 5,
+        "train_loss": 0.2807346156665257,
+        "train_acc": 0.9255952380952381,
+        "lr_head": 4.145780316514581e-05,
+        "lr_backbone": 1.2437340949543742e-05,
+        "val_loss": 0.4557582457860311,
+        "val_accuracy": 0.8,
+        "val_macro_f1": 0.7991071428571428,
+        "val_weighted_f1": 0.7991071428571428
+      },
+      {
+        "epoch": 6,
+        "train_loss": 0.2691161029395603,
+        "train_acc": 0.9375,
+        "lr_head": 3.5531521571796694e-05,
+        "lr_backbone": 1.0659456471539008e-05,
+        "val_loss": 0.44928998152414956,
+        "val_accuracy": 0.8333333333333334,
+        "val_macro_f1": 0.8331479421579533,
+        "val_weighted_f1": 0.8331479421579533
+      },
+      {
+        "epoch": 7,
+        "train_loss": 0.23485895210788363,
+        "train_acc": 0.9375,
+        "lr_head": 2.877229224726381e-05,
+        "lr_backbone": 8.631687674179142e-06,
+        "val_loss": 0.45204214652379354,
+        "val_accuracy": 0.85,
+        "val_macro_f1": 0.849624060150376,
+        "val_weighted_f1": 0.849624060150376
+      },
+      {
+        "epoch": 8,
+        "train_loss": 0.22606890329292842,
+        "train_acc": 0.9553571428571429,
+        "lr_head": 2.1727707752736196e-05,
+        "lr_backbone": 6.518312325820858e-06,
+        "val_loss": 0.4554001530011495,
+        "val_accuracy": 0.7833333333333333,
+        "val_macro_f1": 0.783273131425396,
+        "val_weighted_f1": 0.783273131425396
+      },
+      {
+        "epoch": 9,
+        "train_loss": 0.19964813192685446,
+        "train_acc": 0.9672619047619048,
+        "lr_head": 1.4968478428203314e-05,
+        "lr_backbone": 4.490543528460994e-06,
+        "val_loss": 0.45547234614690146,
+        "val_accuracy": 0.8333333333333334,
+        "val_macro_f1": 0.8325892857142857,
+        "val_weighted_f1": 0.8325892857142857
+      },
+      {
+        "epoch": 10,
+        "train_loss": 0.21034001807371774,
+        "train_acc": 0.9583333333333334,
+        "lr_head": 9.042196834854196e-06,
+        "lr_backbone": 2.712659050456259e-06,
+        "val_loss": 0.45674578348795575,
+        "val_accuracy": 0.8333333333333334,
+        "val_macro_f1": 0.8325892857142857,
+        "val_weighted_f1": 0.8325892857142857
+      },
+      {
+        "epoch": 11,
+        "train_loss": 0.1949910166717711,
+        "train_acc": 0.9702380952380952,
+        "lr_head": 4.428975062428262e-06,
+        "lr_backbone": 1.3286925187284787e-06,
+        "val_loss": 0.4595174193382263,
+        "val_accuracy": 0.8166666666666667,
+        "val_macro_f1": 0.8153846153846154,
+        "val_weighted_f1": 0.8153846153846154
+      },
+      {
+        "epoch": 12,
+        "train_loss": 0.1842061841771716,
+        "train_acc": 0.9761904761904762,
+        "lr_head": 1.502548903041193e-06,
+        "lr_backbone": 4.5076467091235787e-07,
+        "val_loss": 0.4608180999755859,
+        "val_accuracy": 0.8166666666666667,
+        "val_macro_f1": 0.8153846153846154,
+        "val_weighted_f1": 0.8153846153846154
+      }
+    ]
+  },
+  "confusion_matrix": [
+    [
+      50,
+      10
+    ],
+    [
+      13,
+      47
+    ]
+  ],
+  "report": "              precision    recall  f1-score   support\n\n      Gyuyig       0.79      0.83      0.81        60\n     Tsugdri       0.82      0.78      0.80        60\n\n    accuracy                           0.81       120\n   macro avg       0.81      0.81      0.81       120\nweighted avg       0.81      0.81      0.81       120\n",
+  "idx_to_label": {
+    "0": "Gyuyig",
+    "1": "Tsugdri"
+  },
+  "label_to_idx": {
+    "Gyuyig": 0,
+    "Tsugdri": 1
+  },
+  "model_id": "facebook/dinov3-vits16-pretrain-lvd1689m"
+}

split_stats.json

@@ -0,0 +1,35 @@
+{
+  "source": "hf",
+  "classes": [
+    "Gyuyig",
+    "Tsugdri"
+  ],
+  "num_classes": 2,
+  "image_counts_per_class": {
+    "train": {
+      "Gyuyig": 171,
+      "Tsugdri": 171
+    },
+    "val": {
+      "Gyuyig": 30,
+      "Tsugdri": 30
+    },
+    "test": {}
+  },
+  "images_per_split": {
+    "train": 342,
+    "val": 60,
+    "test": 0
+  },
+  "images_total": 402,
+  "image_counts_total_per_class": {
+    "Gyuyig": 201,
+    "Tsugdri": 201
+  },
+  "balanced_parquet_dir": "None",
+  "balanced_dataset_repo": "BDRC/gyuyig-tsugdri-binary-balanced-script-classification-dataset",
+  "task": "gyuyig_tsugdri_binary_classification",
+  "benchmark_dataset_repo": null,
+  "benchmark_per_parent": 60,
+  "test_size": 120
+}

split_stats.md

@@ -0,0 +1,20 @@
+# Split statistics
+
+- **Source:** `hf`
+- **Total images:** 402
+
+
+## Images per split
+
+| Split | Total |
+|-------|------:|
+| train | 342 |
+| val | 60 |
+| test | 0 |
+
+## Images per class (per split)
+
+| Class | train | val | test | **All** |
+|-------|------:|------:|------:|------:|
+| Gyuyig | 171 | 30 | 0 | 201 |
+| Tsugdri | 171 | 30 | 0 | 201 |