Publish two-stage final LoRA adapters and inference guide

README.md

@@ -0,0 +1,235 @@
+---
+license: apache-2.0
+base_model: HuggingFaceTB/SmolVLM2-2.2B-Instruct
+library_name: peft
+pipeline_tag: image-text-to-text
+tags:
+  - smolvlm2
+  - peft
+  - lora
+  - visual-ranking
+  - playing-card
+  - puzzle
+language:
+  - en
+---
+
+# Puker_Judge
+
+`Puker_Judge` 是基于
+[`HuggingFaceTB/SmolVLM2-2.2B-Instruct`](https://huggingface.co/HuggingFaceTB/SmolVLM2-2.2B-Instruct)
+微调的两阶段扑克牌拼接候选判断模型。
+
+本仓库发布的是两个 PEFT LoRA adapter，而不是重复上传两份基础模型：
+
+- `binary_adapter/`：判断一张已经拼好的候选牌面是 `VALID` 还是 `INVALID`。
+- `rank_adapter/`：从同一组碎片产生的 2–4 个几何可行候选中选择图案最连贯的一个。
+- `processor/`：训练时使用的 SmolVLM2 processor 和 tokenizer。
+
+第二阶段由第一阶段权重初始化后继续训练，但推理时不要同时叠加两个
+adapter。做单候选判断时加载 `binary_adapter`，做多候选选择时加载
+`rank_adapter`。
+
+## 能做什么
+
+这两个模型只负责视觉判断，不负责枚举几何拼法，也不直接输出机械臂坐标：
+
+1. CV/几何算法检测碎片并枚举能够填满目标矩形的候选拼法；
+2. 候选只有一个时，可用 `binary_adapter` 判断其图案是否合理；
+3. 候选有 2–4 个时，用 `rank_adapter` 选择最佳候选；
+4. 候选超过4个时，先去重或使用 `binary_adapter` 逐张筛选到 Top-4，再交给
+   `rank_adapter`。
+
+## 安装
+
+BF16：
+
+```bash
+python -m venv .venv
+source .venv/bin/activate
+pip install -r requirements.txt
+```
+
+在 Linux x86-64 NVIDIA GPU 上使用 bitsandbytes INT4：
+
+```bash
+pip install -r requirements-int4.txt
+```
+
+INT4主要建议用于 `rank_adapter`。实测中，rank adapter 在Joker三候选顺序
+轮换上保持了BF16的判断；但 binary adapter 的一个测试集VALID样本在INT4下翻转
+成了INVALID。因此单候选二分类默认应使用BF16，除非已经在自己的数据上重新验证
+INT4准确率。
+
+## 输入图像规范
+
+### 单候选二分类
+
+输入一张透视矫正后的完整候选牌面：
+
+- 推荐规范尺寸：`600×360` 像素，对应物理尺寸比例 `100:60`；
+- 牌面应紧密裁剪，尽量删除桌面、机械臂、阴影和大面积背景；
+- 所有碎片必须位于同一个目标矩形内；
+- 整张牌旋转180°仍视为同一个正确答案；
+- 不要添加候选编号或多选题边框。
+
+### 多候选排序 board
+
+模型实际接收的是一张包含所有选项的 board，而不是多张独立图片。
+
+- board：`1280×820`；
+- 固定2列×2行布局；
+- 选项数必须是 `2、3或4`，不能超过4；
+- 每个候选先规范化为 `600×360`；
+- 标签必须为连续数字 `1, 2, 3, 4`，放在候选图外部；
+- 少于4个选项时，未使用的格子保持空白；
+- 所有候选必须来自同一张牌、同一组碎片，并且在几何上都可行；
+- 各候选应使用相同的裁剪、尺度、背景和成像处理，避免让模型利用无关差异。
+
+`infer_rank.py` 可以接收2–4张候选图并自动生成符合训练格式的 board。
+
+## Prompt
+
+建议保持训练时的英文 prompt，不要随意改写。
+
+### 二分类 prompt
+
+```text
+Judge whether this geometrically assembled playing card has coherent rank, suit, border, portrait, symbols, and continuous artwork. A whole-card 180-degree rotation is valid. Answer VALID or INVALID only.
+```
+
+输出只能是：
+
+```text
+VALID
+```
+
+或：
+
+```text
+INVALID
+```
+
+### 多候选排序 prompt
+
+下面的 `{labels}` 要根据选项数生成，例如3个候选就是 `1, 2, 3`：
+
+```text
+All displayed candidates are geometrically valid reconstructions made from the same playing-card pieces. Select the candidate whose rank, suit, outer border, portrait, symbols, and line artwork form one coherent original playing card. A whole-card 180-degree rotation is equivalent. The available labels are {labels}. Answer with one label only.
+```
+
+输出只能是一个候选编号。
+
+## 使用方法
+
+### 单独判断一个候选的正误
+
+BF16：
+
+```bash
+python infer_binary.py candidate.jpg
+```
+
+INT4 NF4（实验性，binary默认推荐BF16）：
+
+```bash
+python infer_binary.py candidate.jpg --int4
+```
+
+输出示例：
+
+```json
+{
+  "prediction": "VALID",
+  "raw_output": "VALID",
+  "quantization": "int4-nf4"
+}
+```
+
+### 从2–4张候选图中选择
+
+脚本会自动生成 `candidate_board.jpg`：
+
+```bash
+python infer_rank.py \
+  candidate_1.jpg \
+  candidate_2.jpg \
+  candidate_3.jpg \
+  --board-output candidate_board.jpg \
+  --int4
+```
+
+输出示例：
+
+```json
+{
+  "selected_label": 2,
+  "selected_file": "/path/to/candidate_2.jpg",
+  "candidate_count": 3
+}
+```
+
+如果已经自行生成了 board：
+
+```bash
+python infer_rank.py \
+  --board-image candidate_board.jpg \
+  --candidate-count 3 \
+  --int4
+```
+
+## 直接用 Transformers + PEFT 加载
+
+```python
+from pathlib import Path
+
+import torch
+from huggingface_hub import snapshot_download
+from peft import PeftModel
+from transformers import AutoModelForImageTextToText, AutoProcessor
+
+repo_dir = Path(snapshot_download("TuWaveGod/Puker_Judge"))
+processor = AutoProcessor.from_pretrained(repo_dir / "processor")
+base = AutoModelForImageTextToText.from_pretrained(
+    "HuggingFaceTB/SmolVLM2-2.2B-Instruct",
+    torch_dtype=torch.bfloat16,
+).to("cuda")
+model = PeftModel.from_pretrained(
+    base,
+    repo_dir / "rank_adapter",
+).eval()
+```
+
+## 训练信息
+
+- 基础模型：SmolVLM2-2.2B-Instruct；
+- LoRA：rank 16，alpha 32，dropout 0.05；
+- 最大图像长边：1280；
+- 最大文本长度：2048；
+- rank训练场景：18,000；
+- rank验证/测试场景：各1,000；
+- binary训练样本：33,300；
+- binary验证/测试样本：各1,850；
+- 使用52张标准扑克牌资产，不包含大小王；
+- 训练样本包括普通随机、偏心平行四边形、中心双切、全等对称和近似对称困难样本。
+
+## 已做的实拍检查
+
+模型曾在训练集外的 Joker 实拍候选上进行顺序轮换检查。正确候选分别位于
+1、2、3号位置时，rank adapter 的原始生成结果分别为1、2、3。使用
+bitsandbytes INT4 NF4 后结果仍保持一致。该结果只是一组定性检查，不应视为完整
+统计评测。
+
+## 限制
+
+- rank adapter 只训练过最多4个选项，禁止直接构造超过4项的大型 board；
+- binary adapter 对INT4量化更敏感，单候选判断默认使用BF16；
+- board内不存在正确答案时，模型仍会被迫选择一个编号；
+- 透视畸变、反光、遮挡、碎片间隙或候选缩放不一致可能影响判断；
+- 模型不保证识别所有未见过的牌面和印刷风格；
+- 输出是视觉判断，不是几何证明，也不是机械臂控制策略；
+- 建议将正确候选轮换到不同board位置重复2–3次，并对原始生成编号投票，以减小位置偏置。
+
+## License
+
+本模型基于 Apache-2.0 许可的 SmolVLM2-2.2B-Instruct。扑克牌源图和训练数据不包含在本仓库中。

SHA256SUMS

@@ -0,0 +1,2 @@
+6537be03f6fbc4141d3ea86c096ed872e1d5656f5a11c4e3f0d4368ae73ddb73  hf_release/binary_adapter/adapter_model.safetensors
+d2715f57f10feb56f9dc22a79df65f4cad907b77f0f0c2881b110bc92db47197  hf_release/rank_adapter/adapter_model.safetensors

binary_adapter/adapter_config.json

@@ -0,0 +1,108 @@
+{
+  "alpha_pattern": {},
+  "auto_mapping": null,
+  "base_model_name_or_path": "HuggingFaceTB/SmolVLM2-2.2B-Instruct",
+  "bias": "none",
+  "corda_config": null,
+  "eva_config": null,
+  "exclude_modules": null,
+  "fan_in_fan_out": false,
+  "inference_mode": true,
+  "init_lora_weights": true,
+  "layer_replication": null,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "loftq_config": {},
+  "lora_alpha": 32,
+  "lora_bias": false,
+  "lora_dropout": 0.05,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "r": 16,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "text_model.layers.22.self_attn.v_proj",
+    "text_model.layers.3.self_attn.q_proj",
+    "text_model.layers.4.self_attn.q_proj",
+    "text_model.layers.0.self_attn.q_proj",
+    "text_model.layers.12.self_attn.v_proj",
+    "text_model.layers.9.self_attn.k_proj",
+    "text_model.layers.11.self_attn.q_proj",
+    "text_model.layers.15.self_attn.v_proj",
+    "text_model.layers.8.self_attn.k_proj",
+    "text_model.layers.7.self_attn.k_proj",
+    "text_model.layers.18.self_attn.k_proj",
+    "o_proj",
+    "text_model.layers.11.self_attn.k_proj",
+    "text_model.layers.13.self_attn.q_proj",
+    "text_model.layers.16.self_attn.q_proj",
+    "text_model.layers.20.self_attn.q_proj",
+    "text_model.layers.17.self_attn.q_proj",
+    "text_model.layers.21.self_attn.v_proj",
+    "text_model.layers.12.self_attn.k_proj",
+    "text_model.layers.18.self_attn.q_proj",
+    "text_model.layers.10.self_attn.v_proj",
+    "text_model.layers.23.self_attn.k_proj",
+    "text_model.layers.1.self_attn.k_proj",
+    "text_model.layers.17.self_attn.k_proj",
+    "text_model.layers.19.self_attn.k_proj",
+    "text_model.layers.2.self_attn.k_proj",
+    "text_model.layers.6.self_attn.k_proj",
+    "text_model.layers.9.self_attn.q_proj",
+    "text_model.layers.6.self_attn.v_proj",
+    "gate_proj",
+    "text_model.layers.0.self_attn.k_proj",
+    "text_model.layers.16.self_attn.v_proj",
+    "text_model.layers.7.self_attn.v_proj",
+    "text_model.layers.6.self_attn.q_proj",
+    "up_proj",
+    "text_model.layers.23.self_attn.v_proj",
+    "text_model.layers.10.self_attn.k_proj",
+    "text_model.layers.7.self_attn.q_proj",
+    "text_model.layers.11.self_attn.v_proj",
+    "text_model.layers.3.self_attn.v_proj",
+    "text_model.layers.13.self_attn.v_proj",
+    "text_model.layers.21.self_attn.k_proj",
+    "text_model.layers.3.self_attn.k_proj",
+    "text_model.layers.22.self_attn.k_proj",
+    "text_model.layers.15.self_attn.q_proj",
+    "text_model.layers.1.self_attn.v_proj",
+    "text_model.layers.19.self_attn.v_proj",
+    "text_model.layers.2.self_attn.v_proj",
+    "text_model.layers.1.self_attn.q_proj",
+    "text_model.layers.20.self_attn.k_proj",
+    "text_model.layers.15.self_attn.k_proj",
+    "down_proj",
+    "text_model.layers.12.self_attn.q_proj",
+    "text_model.layers.16.self_attn.k_proj",
+    "text_model.layers.0.self_attn.v_proj",
+    "text_model.layers.13.self_attn.k_proj",
+    "text_model.layers.14.self_attn.k_proj",
+    "text_model.layers.21.self_attn.q_proj",
+    "text_model.layers.22.self_attn.q_proj",
+    "text_model.layers.18.self_attn.v_proj",
+    "text_model.layers.19.self_attn.q_proj",
+    "text_model.layers.4.self_attn.k_proj",
+    "text_model.layers.5.self_attn.v_proj",
+    "text_model.layers.14.self_attn.q_proj",
+    "text_model.layers.14.self_attn.v_proj",
+    "text_model.layers.2.self_attn.q_proj",
+    "text_model.layers.23.self_attn.q_proj",
+    "text_model.layers.20.self_attn.v_proj",
+    "text_model.layers.9.self_attn.v_proj",
+    "text_model.layers.8.self_attn.v_proj",
+    "text_model.layers.5.self_attn.q_proj",
+    "text_model.layers.10.self_attn.q_proj",
+    "text_model.layers.8.self_attn.q_proj",
+    "text_model.layers.4.self_attn.v_proj",
+    "text_model.layers.17.self_attn.v_proj",
+    "text_model.layers.5.self_attn.k_proj"
+  ],
+  "task_type": "CAUSAL_LM",
+  "trainable_token_indices": null,
+  "use_dora": false,
+  "use_rslora": false
+}

binary_adapter/adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6537be03f6fbc4141d3ea86c096ed872e1d5656f5a11c4e3f0d4368ae73ddb73
+size 72400072

infer_binary.py

@@ -0,0 +1,73 @@
+#!/usr/bin/env python3
+from __future__ import annotations
+
+import argparse
+import re
+from pathlib import Path
+
+from PIL import Image
+
+from puker_judge_utils import (
+    BASE_MODEL_ID,
+    BINARY_PROMPT,
+    MODEL_REPO_ID,
+    encode_image_prompt,
+    generate_answer,
+    load_adapter,
+    print_json,
+)
+
+
+def parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(
+        description="Judge one assembled playing-card candidate as VALID/INVALID."
+    )
+    parser.add_argument("image", type=Path)
+    parser.add_argument("--repo-id", default=MODEL_REPO_ID)
+    parser.add_argument("--base-model-id", default=BASE_MODEL_ID)
+    parser.add_argument(
+        "--int4",
+        action="store_true",
+        help="Use bitsandbytes NF4 weights with BF16 compute.",
+    )
+    return parser.parse_args()
+
+
+def main() -> None:
+    args = parse_args()
+    model, processor, device, timings = load_adapter(
+        "binary_adapter",
+        repo_id=args.repo_id,
+        base_model_id=args.base_model_id,
+        int4=args.int4,
+    )
+    with Image.open(args.image) as source:
+        image = source.convert("RGB").copy()
+    inputs = encode_image_prompt(
+        processor,
+        image,
+        BINARY_PROMPT,
+        device,
+    )
+    raw_output, generation_seconds = generate_answer(
+        model,
+        processor,
+        inputs,
+    )
+    match = re.search(r"\b(INVALID|VALID)\b", raw_output.upper())
+    if match is None:
+        raise SystemExit(f"Model returned an invalid answer: {raw_output!r}")
+    print_json(
+        {
+            "prediction": match.group(1),
+            "raw_output": raw_output,
+            "image": str(args.image.resolve()),
+            "quantization": "int4-nf4" if args.int4 else "bf16",
+            "generation_seconds": round(generation_seconds, 4),
+            **{key: round(value, 4) for key, value in timings.items()},
+        }
+    )
+
+
+if __name__ == "__main__":
+    main()

infer_rank.py

@@ -0,0 +1,206 @@
+#!/usr/bin/env python3
+from __future__ import annotations
+
+import argparse
+import re
+from pathlib import Path
+
+from PIL import Image, ImageDraw, ImageFont
+
+from puker_judge_utils import (
+    BASE_MODEL_ID,
+    MODEL_REPO_ID,
+    encode_image_prompt,
+    generate_answer,
+    load_adapter,
+    print_json,
+    rank_prompt,
+)
+
+BOARD_SIZE = (1280, 820)
+CANDIDATE_SIZE = (600, 360)
+
+
+def parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(
+        description="Choose the best of two to four playing-card candidates."
+    )
+    parser.add_argument(
+        "candidates",
+        type=Path,
+        nargs="*",
+        help="Two to four rectified candidate images.",
+    )
+    parser.add_argument(
+        "--board-image",
+        type=Path,
+        help="Use an already constructed board instead of candidate files.",
+    )
+    parser.add_argument(
+        "--candidate-count",
+        type=int,
+        help="Required with --board-image; must be between 2 and 4.",
+    )
+    parser.add_argument(
+        "--board-output",
+        type=Path,
+        default=Path("candidate_board.jpg"),
+    )
+    parser.add_argument("--repo-id", default=MODEL_REPO_ID)
+    parser.add_argument("--base-model-id", default=BASE_MODEL_ID)
+    parser.add_argument(
+        "--int4",
+        action="store_true",
+        help="Use bitsandbytes NF4 weights with BF16 compute.",
+    )
+    return parser.parse_args()
+
+
+def load_font(size: int) -> ImageFont.ImageFont:
+    for path in (
+        "/usr/share/fonts/truetype/dejavu/DejaVuSans-Bold.ttf",
+        "DejaVuSans-Bold.ttf",
+    ):
+        try:
+            return ImageFont.truetype(path, size=size)
+        except OSError:
+            continue
+    return ImageFont.load_default()
+
+
+def build_board(candidate_paths: list[Path]) -> Image.Image:
+    count = len(candidate_paths)
+    if not 2 <= count <= 4:
+        raise ValueError("Provide exactly 2, 3, or 4 candidate images.")
+
+    board = Image.new("RGB", BOARD_SIZE, (226, 231, 237))
+    draw = ImageDraw.Draw(board)
+    font = load_font(34)
+    outer_margin = 28
+    column_gap = 24
+    row_gap = 24
+    cell_width = (BOARD_SIZE[0] - 2 * outer_margin - column_gap) // 2
+    cell_height = (BOARD_SIZE[1] - 2 * outer_margin - row_gap) // 2
+    label_height = 46
+
+    for index, path in enumerate(candidate_paths):
+        column = index % 2
+        row = index // 2
+        x0 = outer_margin + column * (cell_width + column_gap)
+        y0 = outer_margin + row * (cell_height + row_gap)
+        x1 = x0 + cell_width
+        y1 = y0 + cell_height
+        draw.rounded_rectangle(
+            (x0, y0, x1, y1),
+            radius=14,
+            fill=(244, 246, 248),
+            outline=(178, 184, 192),
+            width=2,
+        )
+        draw.rounded_rectangle(
+            (x0 + 12, y0 + 8, x0 + 76, y0 + label_height),
+            radius=10,
+            fill=(255, 218, 72),
+            outline=(45, 48, 52),
+            width=2,
+        )
+        draw.text(
+            (x0 + 44, y0 + 8 + label_height // 2),
+            str(index + 1),
+            font=font,
+            fill=(25, 31, 42),
+            anchor="mm",
+        )
+        with Image.open(path) as source:
+            candidate = source.convert("RGB").resize(
+                CANDIDATE_SIZE,
+                Image.Resampling.LANCZOS,
+            )
+        available_width = cell_width - 36
+        available_height = cell_height - label_height - 28
+        candidate.thumbnail(
+            (available_width, available_height),
+            Image.Resampling.LANCZOS,
+        )
+        paste_x = x0 + (cell_width - candidate.width) // 2
+        paste_y = y0 + label_height + (
+            cell_height - label_height - candidate.height
+        ) // 2
+        draw.rectangle(
+            (
+                paste_x + 4,
+                paste_y + 5,
+                paste_x + 4 + candidate.width,
+                paste_y + 5 + candidate.height,
+            ),
+            fill=(185, 190, 196),
+        )
+        board.paste(candidate, (paste_x, paste_y))
+    return board
+
+
+def main() -> None:
+    args = parse_args()
+    if args.board_image is not None:
+        if args.candidates:
+            raise SystemExit(
+                "Use either candidate files or --board-image, not both."
+            )
+        if args.candidate_count is None or not 2 <= args.candidate_count <= 4:
+            raise SystemExit(
+                "--candidate-count 2..4 is required with --board-image."
+            )
+        with Image.open(args.board_image) as source:
+            board = source.convert("RGB").copy()
+        candidate_count = args.candidate_count
+        candidate_paths: list[Path] = []
+    else:
+        if args.candidate_count is not None:
+            raise SystemExit(
+                "--candidate-count is inferred when candidate files are used."
+            )
+        candidate_paths = args.candidates
+        candidate_count = len(candidate_paths)
+        board = build_board(candidate_paths)
+        args.board_output.parent.mkdir(parents=True, exist_ok=True)
+        board.save(args.board_output, "JPEG", quality=96, subsampling=0)
+
+    model, processor, device, timings = load_adapter(
+        "rank_adapter",
+        repo_id=args.repo_id,
+        base_model_id=args.base_model_id,
+        int4=args.int4,
+    )
+    inputs = encode_image_prompt(
+        processor,
+        board,
+        rank_prompt(candidate_count),
+        device,
+    )
+    raw_output, generation_seconds = generate_answer(
+        model,
+        processor,
+        inputs,
+    )
+    match = re.search(r"[1-4]", raw_output)
+    if match is None or int(match.group(0)) > candidate_count:
+        raise SystemExit(f"Model returned an invalid answer: {raw_output!r}")
+    label = int(match.group(0))
+    payload = {
+        "selected_label": label,
+        "raw_output": raw_output,
+        "candidate_count": candidate_count,
+        "quantization": "int4-nf4" if args.int4 else "bf16",
+        "generation_seconds": round(generation_seconds, 4),
+        **{key: round(value, 4) for key, value in timings.items()},
+    }
+    if candidate_paths:
+        payload["selected_file"] = str(candidate_paths[label - 1].resolve())
+        payload["board_image"] = str(args.board_output.resolve())
+    else:
+        payload["board_image"] = str(args.board_image.resolve())
+    print_json(payload)
+
+
+if __name__ == "__main__":
+    main()

processor/added_tokens.json

@@ -0,0 +1,130 @@
+{
+  "<end_of_utterance>": 49279,
+  "<fake_token_around_image>": 49189,
+  "<global-img>": 49152,
+  "<image>": 49190,
+  "<row_1_col_1>": 49153,
+  "<row_1_col_2>": 49154,
+  "<row_1_col_3>": 49155,
+  "<row_1_col_4>": 49156,
+  "<row_1_col_5>": 49157,
+  "<row_1_col_6>": 49158,
+  "<row_2_col_1>": 49159,
+  "<row_2_col_2>": 49160,
+  "<row_2_col_3>": 49161,
+  "<row_2_col_4>": 49162,
+  "<row_2_col_5>": 49163,
+  "<row_2_col_6>": 49164,
+  "<row_3_col_1>": 49165,
+  "<row_3_col_2>": 49166,
+  "<row_3_col_3>": 49167,
+  "<row_3_col_4>": 49168,
+  "<row_3_col_5>": 49169,
+  "<row_3_col_6>": 49170,
+  "<row_4_col_1>": 49171,
+  "<row_4_col_2>": 49172,
+  "<row_4_col_3>": 49173,
+  "<row_4_col_4>": 49174,
+  "<row_4_col_5>": 49175,
+  "<row_4_col_6>": 49176,
+  "<row_5_col_1>": 49177,
+  "<row_5_col_2>": 49178,
+  "<row_5_col_3>": 49179,
+  "<row_5_col_4>": 49180,
+  "<row_5_col_5>": 49181,
+  "<row_5_col_6>": 49182,
+  "<row_6_col_1>": 49183,
+  "<row_6_col_2>": 49184,
+  "<row_6_col_3>": 49185,
+  "<row_6_col_4>": 49186,
+  "<row_6_col_5>": 49187,
+  "<row_6_col_6>": 49188,
+  "<|reserved_special_token_0|>": 49191,
+  "<|reserved_special_token_10|>": 49201,
+  "<|reserved_special_token_11|>": 49202,
+  "<|reserved_special_token_12|>": 49203,
+  "<|reserved_special_token_13|>": 49204,
+  "<|reserved_special_token_14|>": 49205,
+  "<|reserved_special_token_15|>": 49206,
+  "<|reserved_special_token_16|>": 49207,
+  "<|reserved_special_token_17|>": 49208,
+  "<|reserved_special_token_18|>": 49209,
+  "<|reserved_special_token_19|>": 49210,
+  "<|reserved_special_token_1|>": 49192,
+  "<|reserved_special_token_20|>": 49211,
+  "<|reserved_special_token_21|>": 49212,
+  "<|reserved_special_token_22|>": 49213,
+  "<|reserved_special_token_23|>": 49214,
+  "<|reserved_special_token_24|>": 49215,
+  "<|reserved_special_token_25|>": 49216,
+  "<|reserved_special_token_26|>": 49217,
+  "<|reserved_special_token_27|>": 49218,
+  "<|reserved_special_token_28|>": 49219,
+  "<|reserved_special_token_29|>": 49220,
+  "<|reserved_special_token_2|>": 49193,
+  "<|reserved_special_token_30|>": 49221,
+  "<|reserved_special_token_31|>": 49222,
+  "<|reserved_special_token_32|>": 49223,
+  "<|reserved_special_token_33|>": 49224,
+  "<|reserved_special_token_34|>": 49225,
+  "<|reserved_special_token_35|>": 49226,
+  "<|reserved_special_token_36|>": 49227,
+  "<|reserved_special_token_37|>": 49228,
+  "<|reserved_special_token_38|>": 49229,
+  "<|reserved_special_token_39|>": 49230,
+  "<|reserved_special_token_3|>": 49194,
+  "<|reserved_special_token_40|>": 49231,
+  "<|reserved_special_token_41|>": 49232,
+  "<|reserved_special_token_42|>": 49233,
+  "<|reserved_special_token_43|>": 49234,
+  "<|reserved_special_token_44|>": 49235,
+  "<|reserved_special_token_45|>": 49236,
+  "<|reserved_special_token_46|>": 49237,
+  "<|reserved_special_token_47|>": 49238,
+  "<|reserved_special_token_48|>": 49239,
+  "<|reserved_special_token_49|>": 49240,
+  "<|reserved_special_token_4|>": 49195,
+  "<|reserved_special_token_50|>": 49241,
+  "<|reserved_special_token_51|>": 49242,
+  "<|reserved_special_token_52|>": 49243,
+  "<|reserved_special_token_53|>": 49244,
+  "<|reserved_special_token_54|>": 49245,
+  "<|reserved_special_token_55|>": 49246,
+  "<|reserved_special_token_56|>": 49247,
+  "<|reserved_special_token_57|>": 49248,
+  "<|reserved_special_token_58|>": 49249,
+  "<|reserved_special_token_59|>": 49250,
+  "<|reserved_special_token_5|>": 49196,
+  "<|reserved_special_token_60|>": 49251,
+  "<|reserved_special_token_61|>": 49252,
+  "<|reserved_special_token_62|>": 49253,
+  "<|reserved_special_token_63|>": 49254,
+  "<|reserved_special_token_64|>": 49255,
+  "<|reserved_special_token_65|>": 49256,
+  "<|reserved_special_token_66|>": 49257,
+  "<|reserved_special_token_67|>": 49258,
+  "<|reserved_special_token_68|>": 49259,
+  "<|reserved_special_token_69|>": 49260,
+  "<|reserved_special_token_6|>": 49197,
+  "<|reserved_special_token_70|>": 49261,
+  "<|reserved_special_token_71|>": 49262,
+  "<|reserved_special_token_72|>": 49263,
+  "<|reserved_special_token_73|>": 49264,
+  "<|reserved_special_token_74|>": 49265,
+  "<|reserved_special_token_75|>": 49266,
+  "<|reserved_special_token_76|>": 49267,
+  "<|reserved_special_token_77|>": 49268,
+  "<|reserved_special_token_78|>": 49269,
+  "<|reserved_special_token_79|>": 49270,
+  "<|reserved_special_token_7|>": 49198,
+  "<|reserved_special_token_80|>": 49271,
+  "<|reserved_special_token_81|>": 49272,
+  "<|reserved_special_token_82|>": 49273,
+  "<|reserved_special_token_83|>": 49274,
+  "<|reserved_special_token_84|>": 49275,
+  "<|reserved_special_token_85|>": 49276,
+  "<|reserved_special_token_86|>": 49277,
+  "<|reserved_special_token_87|>": 49278,
+  "<|reserved_special_token_8|>": 49199,
+  "<|reserved_special_token_9|>": 49200
+}

processor/chat_template.jinja

@@ -0,0 +1,2 @@
+<|im_start|>{% for message in messages %}{{message['role'] | capitalize}}{% if message['content'][0]['type'] == 'image' %}{{':'}}{% else %}{{': '}}{% endif %}{% for line in message['content'] %}{% if line['type'] == 'text' %}{{line['text']}}{% elif line['type'] == 'image' %}{{ '<image>' }}{% endif %}{% endfor %}<end_of_utterance>
+{% endfor %}{% if add_generation_prompt %}{{ 'Assistant:' }}{% endif %}

processor/preprocessor_config.json

@@ -0,0 +1,35 @@
+{
+  "do_convert_rgb": true,
+  "do_image_splitting": true,
+  "do_normalize": true,
+  "do_pad": true,
+  "do_rescale": true,
+  "do_resize": true,
+  "image_mean": [
+    0.5,
+    0.5,
+    0.5
+  ],
+  "image_processor_type": "SmolVLMImageProcessor",
+  "image_std": [
+    0.5,
+    0.5,
+    0.5
+  ],
+  "max_image_size": {
+    "longest_edge": 384
+  },
+  "processor_class": "SmolVLMProcessor",
+  "resample": 1,
+  "rescale_factor": 0.00392156862745098,
+  "size": {
+    "longest_edge": 1536
+  },
+  "video_sampling": {
+    "fps": 1,
+    "max_frames": 64,
+    "video_size": {
+      "longest_edge": 384
+    }
+  }
+}

processor/processor_config.json

@@ -0,0 +1,4 @@
+{
+  "image_seq_len": 81,
+  "processor_class": "SmolVLMProcessor"
+}

processor/special_tokens_map.json

@@ -0,0 +1,39 @@
+{
+  "additional_special_tokens": [
+    "<fake_token_around_image>",
+    "<image>",
+    "<end_of_utterance>"
+  ],
+  "bos_token": {
+    "content": "<|im_start|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "end_of_utterance_token": "<end_of_utterance>",
+  "eos_token": {
+    "content": "<end_of_utterance>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "fake_image_token": "<fake_token_around_image>",
+  "global_image_token": "<global-img>",
+  "image_token": "<image>",
+  "pad_token": {
+    "content": "<|im_end|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "unk_token": {
+    "content": "<|endoftext|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  }
+}

processor/tokenizer_config.json

@@ -0,0 +1,1191 @@
+{
+  "add_prefix_space": false,
+  "added_tokens_decoder": {
+    "0": {
+      "content": "<|endoftext|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "1": {
+      "content": "<|im_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "2": {
+      "content": "<|im_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "3": {
+      "content": "<repo_name>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "4": {
+      "content": "<reponame>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "5": {
+      "content": "<file_sep>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "6": {
+      "content": "<filename>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "7": {
+      "content": "<gh_stars>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "8": {
+      "content": "<issue_start>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "9": {
+      "content": "<issue_comment>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "10": {
+      "content": "<issue_closed>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "11": {
+      "content": "<jupyter_start>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "12": {
+      "content": "<jupyter_text>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "13": {
+      "content": "<jupyter_code>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "14": {
+      "content": "<jupyter_output>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "15": {
+      "content": "<jupyter_script>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "16": {
+      "content": "<empty_output>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49152": {
+      "content": "<global-img>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49153": {
+      "content": "<row_1_col_1>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49154": {
+      "content": "<row_1_col_2>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49155": {
+      "content": "<row_1_col_3>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49156": {
+      "content": "<row_1_col_4>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49157": {
+      "content": "<row_1_col_5>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49158": {
+      "content": "<row_1_col_6>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49159": {
+      "content": "<row_2_col_1>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49160": {
+      "content": "<row_2_col_2>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49161": {
+      "content": "<row_2_col_3>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49162": {
+      "content": "<row_2_col_4>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49163": {
+      "content": "<row_2_col_5>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49164": {
+      "content": "<row_2_col_6>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49165": {
+      "content": "<row_3_col_1>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49166": {
+      "content": "<row_3_col_2>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49167": {
+      "content": "<row_3_col_3>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49168": {
+      "content": "<row_3_col_4>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49169": {
+      "content": "<row_3_col_5>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49170": {
+      "content": "<row_3_col_6>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49171": {
+      "content": "<row_4_col_1>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49172": {
+      "content": "<row_4_col_2>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49173": {
+      "content": "<row_4_col_3>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49174": {
+      "content": "<row_4_col_4>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49175": {
+      "content": "<row_4_col_5>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49176": {
+      "content": "<row_4_col_6>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49177": {
+      "content": "<row_5_col_1>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49178": {
+      "content": "<row_5_col_2>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49179": {
+      "content": "<row_5_col_3>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49180": {
+      "content": "<row_5_col_4>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49181": {
+      "content": "<row_5_col_5>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49182": {
+      "content": "<row_5_col_6>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49183": {
+      "content": "<row_6_col_1>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49184": {
+      "content": "<row_6_col_2>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49185": {
+      "content": "<row_6_col_3>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49186": {
+      "content": "<row_6_col_4>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49187": {
+      "content": "<row_6_col_5>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49188": {
+      "content": "<row_6_col_6>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49189": {
+      "content": "<fake_token_around_image>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49190": {
+      "content": "<image>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49191": {
+      "content": "<|reserved_special_token_0|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49192": {
+      "content": "<|reserved_special_token_1|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49193": {
+      "content": "<|reserved_special_token_2|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49194": {
+      "content": "<|reserved_special_token_3|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49195": {
+      "content": "<|reserved_special_token_4|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49196": {
+      "content": "<|reserved_special_token_5|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49197": {
+      "content": "<|reserved_special_token_6|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49198": {
+      "content": "<|reserved_special_token_7|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49199": {
+      "content": "<|reserved_special_token_8|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49200": {
+      "content": "<|reserved_special_token_9|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49201": {
+      "content": "<|reserved_special_token_10|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49202": {
+      "content": "<|reserved_special_token_11|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49203": {
+      "content": "<|reserved_special_token_12|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49204": {
+      "content": "<|reserved_special_token_13|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49205": {
+      "content": "<|reserved_special_token_14|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49206": {
+      "content": "<|reserved_special_token_15|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49207": {
+      "content": "<|reserved_special_token_16|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49208": {
+      "content": "<|reserved_special_token_17|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49209": {
+      "content": "<|reserved_special_token_18|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49210": {
+      "content": "<|reserved_special_token_19|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49211": {
+      "content": "<|reserved_special_token_20|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49212": {
+      "content": "<|reserved_special_token_21|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49213": {
+      "content": "<|reserved_special_token_22|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49214": {
+      "content": "<|reserved_special_token_23|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49215": {
+      "content": "<|reserved_special_token_24|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49216": {
+      "content": "<|reserved_special_token_25|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49217": {
+      "content": "<|reserved_special_token_26|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49218": {
+      "content": "<|reserved_special_token_27|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49219": {
+      "content": "<|reserved_special_token_28|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49220": {
+      "content": "<|reserved_special_token_29|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49221": {
+      "content": "<|reserved_special_token_30|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49222": {
+      "content": "<|reserved_special_token_31|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49223": {
+      "content": "<|reserved_special_token_32|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49224": {
+      "content": "<|reserved_special_token_33|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49225": {
+      "content": "<|reserved_special_token_34|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49226": {
+      "content": "<|reserved_special_token_35|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49227": {
+      "content": "<|reserved_special_token_36|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49228": {
+      "content": "<|reserved_special_token_37|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49229": {
+      "content": "<|reserved_special_token_38|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49230": {
+      "content": "<|reserved_special_token_39|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49231": {
+      "content": "<|reserved_special_token_40|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49232": {
+      "content": "<|reserved_special_token_41|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49233": {
+      "content": "<|reserved_special_token_42|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49234": {
+      "content": "<|reserved_special_token_43|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49235": {
+      "content": "<|reserved_special_token_44|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49236": {
+      "content": "<|reserved_special_token_45|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49237": {
+      "content": "<|reserved_special_token_46|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49238": {
+      "content": "<|reserved_special_token_47|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49239": {
+      "content": "<|reserved_special_token_48|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49240": {
+      "content": "<|reserved_special_token_49|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49241": {
+      "content": "<|reserved_special_token_50|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49242": {
+      "content": "<|reserved_special_token_51|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49243": {
+      "content": "<|reserved_special_token_52|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49244": {
+      "content": "<|reserved_special_token_53|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49245": {
+      "content": "<|reserved_special_token_54|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49246": {
+      "content": "<|reserved_special_token_55|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49247": {
+      "content": "<|reserved_special_token_56|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49248": {
+      "content": "<|reserved_special_token_57|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49249": {
+      "content": "<|reserved_special_token_58|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49250": {
+      "content": "<|reserved_special_token_59|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49251": {
+      "content": "<|reserved_special_token_60|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49252": {
+      "content": "<|reserved_special_token_61|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49253": {
+      "content": "<|reserved_special_token_62|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49254": {
+      "content": "<|reserved_special_token_63|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49255": {
+      "content": "<|reserved_special_token_64|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49256": {
+      "content": "<|reserved_special_token_65|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49257": {
+      "content": "<|reserved_special_token_66|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49258": {
+      "content": "<|reserved_special_token_67|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49259": {
+      "content": "<|reserved_special_token_68|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49260": {
+      "content": "<|reserved_special_token_69|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49261": {
+      "content": "<|reserved_special_token_70|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49262": {
+      "content": "<|reserved_special_token_71|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49263": {
+      "content": "<|reserved_special_token_72|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49264": {
+      "content": "<|reserved_special_token_73|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49265": {
+      "content": "<|reserved_special_token_74|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49266": {
+      "content": "<|reserved_special_token_75|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49267": {
+      "content": "<|reserved_special_token_76|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49268": {
+      "content": "<|reserved_special_token_77|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49269": {
+      "content": "<|reserved_special_token_78|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49270": {
+      "content": "<|reserved_special_token_79|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49271": {
+      "content": "<|reserved_special_token_80|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49272": {
+      "content": "<|reserved_special_token_81|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49273": {
+      "content": "<|reserved_special_token_82|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49274": {
+      "content": "<|reserved_special_token_83|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49275": {
+      "content": "<|reserved_special_token_84|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49276": {
+      "content": "<|reserved_special_token_85|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49277": {
+      "content": "<|reserved_special_token_86|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49278": {
+      "content": "<|reserved_special_token_87|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "49279": {
+      "content": "<end_of_utterance>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    }
+  },
+  "additional_special_tokens": [
+    "<fake_token_around_image>",
+    "<image>",
+    "<end_of_utterance>"
+  ],
+  "bos_token": "<|im_start|>",
+  "clean_up_tokenization_spaces": false,
+  "end_of_utterance_token": "<end_of_utterance>",
+  "eos_token": "<end_of_utterance>",
+  "extra_special_tokens": {
+    "end_of_utterance_token": "<end_of_utterance>",
+    "fake_image_token": "<fake_token_around_image>",
+    "global_image_token": "<global-img>",
+    "image_token": "<image>"
+  },
+  "fake_image_token": "<fake_token_around_image>",
+  "global_image_token": "<global-img>",
+  "image_token": "<image>",
+  "legacy": false,
+  "model_max_length": 16384,
+  "pad_token": "<|im_end|>",
+  "processor_class": "SmolVLMProcessor",
+  "tokenizer_class": "GPT2Tokenizer",
+  "truncation_side": "left",
+  "unk_token": "<|endoftext|>",
+  "vocab_size": 49152
+}

puker_judge_utils.py

@@ -0,0 +1,193 @@
+from __future__ import annotations
+
+import json
+import time
+from pathlib import Path
+from typing import Any
+
+import torch
+from huggingface_hub import snapshot_download
+from peft import PeftModel
+from PIL import Image
+from transformers import (
+    AutoModelForImageTextToText,
+    AutoProcessor,
+    BitsAndBytesConfig,
+)
+
+BASE_MODEL_ID = "HuggingFaceTB/SmolVLM2-2.2B-Instruct"
+MODEL_REPO_ID = "TuWaveGod/Puker_Judge"
+MAX_LENGTH = 2048
+MAX_IMAGE_LONGEST_EDGE = 1280
+
+BINARY_PROMPT = (
+    "Judge whether this geometrically assembled playing card has coherent rank, "
+    "suit, border, portrait, symbols, and continuous artwork. A whole-card "
+    "180-degree rotation is valid. Answer VALID or INVALID only."
+)
+
+
+def rank_prompt(candidate_count: int) -> str:
+    if not 2 <= candidate_count <= 4:
+        raise ValueError("Rank inference requires 2 to 4 candidates.")
+    labels = ", ".join(str(index) for index in range(1, candidate_count + 1))
+    return (
+        "All displayed candidates are geometrically valid reconstructions made "
+        "from the same playing-card pieces. Select the candidate whose rank, suit, "
+        "outer border, portrait, symbols, and line artwork form one coherent "
+        "original playing card. A whole-card 180-degree rotation is equivalent. "
+        f"The available labels are {labels}. Answer with one label only."
+    )
+
+
+def resize_for_model(image: Image.Image) -> Image.Image:
+    image = image.convert("RGB")
+    longest = max(image.size)
+    if longest <= MAX_IMAGE_LONGEST_EDGE:
+        return image
+    scale = MAX_IMAGE_LONGEST_EDGE / longest
+    return image.resize(
+        (
+            max(1, int(round(image.width * scale))),
+            max(1, int(round(image.height * scale))),
+        ),
+        Image.Resampling.LANCZOS,
+    )
+
+
+def load_adapter(
+    adapter_name: str,
+    *,
+    repo_id: str = MODEL_REPO_ID,
+    base_model_id: str = BASE_MODEL_ID,
+    int4: bool = False,
+) -> tuple[Any, Any, torch.device, dict[str, float]]:
+    if adapter_name not in {"binary_adapter", "rank_adapter"}:
+        raise ValueError(f"Unknown adapter: {adapter_name}")
+    if not torch.cuda.is_available():
+        raise RuntimeError("A CUDA GPU is required by these example scripts.")
+
+    local_repo = Path(repo_id).expanduser()
+    if local_repo.is_dir():
+        snapshot_path = local_repo.resolve()
+        download_seconds = 0.0
+    else:
+        download_started = time.perf_counter()
+        snapshot_path = Path(
+            snapshot_download(
+                repo_id=repo_id,
+                allow_patterns=[
+                    f"{adapter_name}/*",
+                    "processor/*",
+                ],
+            )
+        )
+        download_seconds = time.perf_counter() - download_started
+
+    processor = AutoProcessor.from_pretrained(snapshot_path / "processor")
+    load_kwargs: dict[str, Any] = {
+        "torch_dtype": torch.bfloat16,
+        "attn_implementation": "sdpa",
+    }
+    if int4:
+        load_kwargs.update(
+            {
+                "quantization_config": BitsAndBytesConfig(
+                    load_in_4bit=True,
+                    bnb_4bit_quant_type="nf4",
+                    bnb_4bit_compute_dtype=torch.bfloat16,
+                    bnb_4bit_use_double_quant=True,
+                ),
+                "device_map": {"": 0},
+            }
+        )
+
+    load_started = time.perf_counter()
+    base_model = AutoModelForImageTextToText.from_pretrained(
+        base_model_id,
+        **load_kwargs,
+    )
+    if not int4:
+        base_model = base_model.to("cuda:0")
+    model = PeftModel.from_pretrained(
+        base_model,
+        snapshot_path / adapter_name,
+    ).eval()
+    torch.cuda.synchronize()
+    load_seconds = time.perf_counter() - load_started
+    return (
+        model,
+        processor,
+        torch.device("cuda:0"),
+        {
+            "snapshot_download_seconds": download_seconds,
+            "model_load_seconds": load_seconds,
+        },
+    )
+
+
+def encode_image_prompt(
+    processor: Any,
+    image: Image.Image,
+    prompt: str,
+    device: torch.device,
+) -> dict[str, Any]:
+    messages = [
+        {
+            "role": "user",
+            "content": [
+                {"type": "image"},
+                {"type": "text", "text": prompt},
+            ],
+        }
+    ]
+    text = processor.apply_chat_template(
+        messages,
+        add_generation_prompt=True,
+        tokenize=False,
+    )
+    inputs = processor(
+        text=text,
+        images=resize_for_model(image),
+        return_tensors="pt",
+        truncation=True,
+        max_length=MAX_LENGTH,
+    )
+    moved: dict[str, Any] = {}
+    for key, value in inputs.items():
+        if not isinstance(value, torch.Tensor):
+            moved[key] = value
+        elif key == "pixel_values":
+            moved[key] = value.to(device=device, dtype=torch.bfloat16)
+        else:
+            moved[key] = value.to(device=device)
+    return moved
+
+
+def generate_answer(
+    model: Any,
+    processor: Any,
+    inputs: dict[str, Any],
+    *,
+    max_new_tokens: int = 4,
+) -> tuple[str, float]:
+    input_length = int(inputs["input_ids"].shape[1])
+    torch.cuda.synchronize()
+    started = time.perf_counter()
+    with torch.inference_mode():
+        output_ids = model.generate(
+            **inputs,
+            do_sample=False,
+            max_new_tokens=max_new_tokens,
+        )
+    torch.cuda.synchronize()
+    elapsed = time.perf_counter() - started
+    answer = processor.decode(
+        output_ids[0, input_length:],
+        skip_special_tokens=True,
+    ).strip()
+    return answer, elapsed
+
+
+def print_json(payload: dict[str, Any]) -> None:
+    print(json.dumps(payload, ensure_ascii=False, indent=2))

rank_adapter/adapter_config.json

@@ -0,0 +1,108 @@
+{
+  "alpha_pattern": {},
+  "auto_mapping": null,
+  "base_model_name_or_path": "HuggingFaceTB/SmolVLM2-2.2B-Instruct",
+  "bias": "none",
+  "corda_config": null,
+  "eva_config": null,
+  "exclude_modules": null,
+  "fan_in_fan_out": false,
+  "inference_mode": true,
+  "init_lora_weights": true,
+  "layer_replication": null,
+  "layers_pattern": null,
+  "layers_to_transform": null,
+  "loftq_config": {},
+  "lora_alpha": 32,
+  "lora_bias": false,
+  "lora_dropout": 0.05,
+  "megatron_config": null,
+  "megatron_core": "megatron.core",
+  "modules_to_save": null,
+  "peft_type": "LORA",
+  "r": 16,
+  "rank_pattern": {},
+  "revision": null,
+  "target_modules": [
+    "text_model.layers.20.self_attn.k_proj",
+    "text_model.layers.21.self_attn.k_proj",
+    "text_model.layers.23.self_attn.v_proj",
+    "text_model.layers.20.self_attn.q_proj",
+    "text_model.layers.17.self_attn.q_proj",
+    "text_model.layers.7.self_attn.v_proj",
+    "text_model.layers.3.self_attn.v_proj",
+    "text_model.layers.19.self_attn.v_proj",
+    "text_model.layers.2.self_attn.v_proj",
+    "text_model.layers.0.self_attn.v_proj",
+    "text_model.layers.21.self_attn.q_proj",
+    "text_model.layers.5.self_attn.q_proj",
+    "text_model.layers.12.self_attn.v_proj",
+    "text_model.layers.23.self_attn.k_proj",
+    "text_model.layers.16.self_attn.v_proj",
+    "text_model.layers.7.self_attn.q_proj",
+    "text_model.layers.4.self_attn.q_proj",
+    "o_proj",
+    "text_model.layers.8.self_attn.k_proj",
+    "text_model.layers.0.self_attn.q_proj",
+    "text_model.layers.9.self_attn.k_proj",
+    "text_model.layers.6.self_attn.v_proj",
+    "text_model.layers.23.self_attn.q_proj",
+    "text_model.layers.19.self_attn.q_proj",
+    "text_model.layers.8.self_attn.v_proj",
+    "text_model.layers.15.self_attn.k_proj",
+    "text_model.layers.15.self_attn.v_proj",
+    "up_proj",
+    "text_model.layers.9.self_attn.q_proj",
+    "text_model.layers.1.self_attn.q_proj",
+    "text_model.layers.4.self_attn.v_proj",
+    "text_model.layers.11.self_attn.v_proj",
+    "text_model.layers.9.self_attn.v_proj",
+    "text_model.layers.14.self_attn.k_proj",
+    "text_model.layers.17.self_attn.k_proj",
+    "text_model.layers.4.self_attn.k_proj",
+    "text_model.layers.2.self_attn.q_proj",
+    "text_model.layers.16.self_attn.k_proj",
+    "text_model.layers.14.self_attn.q_proj",
+    "text_model.layers.22.self_attn.q_proj",
+    "text_model.layers.5.self_attn.k_proj",
+    "text_model.layers.10.self_attn.v_proj",
+    "text_model.layers.15.self_attn.q_proj",
+    "text_model.layers.13.self_attn.q_proj",
+    "text_model.layers.0.self_attn.k_proj",
+    "down_proj",
+    "text_model.layers.19.self_attn.k_proj",
+    "text_model.layers.6.self_attn.q_proj",
+    "text_model.layers.10.self_attn.k_proj",
+    "text_model.layers.22.self_attn.v_proj",
+    "text_model.layers.2.self_attn.k_proj",
+    "text_model.layers.20.self_attn.v_proj",
+    "text_model.layers.13.self_attn.v_proj",
+    "text_model.layers.5.self_attn.v_proj",
+    "text_model.layers.6.self_attn.k_proj",
+    "text_model.layers.12.self_attn.q_proj",
+    "text_model.layers.21.self_attn.v_proj",
+    "text_model.layers.14.self_attn.v_proj",
+    "text_model.layers.16.self_attn.q_proj",
+    "text_model.layers.3.self_attn.q_proj",
+    "text_model.layers.18.self_attn.v_proj",
+    "text_model.layers.11.self_attn.k_proj",
+    "gate_proj",
+    "text_model.layers.11.self_attn.q_proj",
+    "text_model.layers.7.self_attn.k_proj",
+    "text_model.layers.12.self_attn.k_proj",
+    "text_model.layers.18.self_attn.q_proj",
+    "text_model.layers.1.self_attn.k_proj",
+    "text_model.layers.3.self_attn.k_proj",
+    "text_model.layers.22.self_attn.k_proj",
+    "text_model.layers.1.self_attn.v_proj",
+    "text_model.layers.8.self_attn.q_proj",
+    "text_model.layers.17.self_attn.v_proj",
+    "text_model.layers.13.self_attn.k_proj",
+    "text_model.layers.10.self_attn.q_proj",
+    "text_model.layers.18.self_attn.k_proj"
+  ],
+  "task_type": "CAUSAL_LM",
+  "trainable_token_indices": null,
+  "use_dora": false,
+  "use_rslora": false
+}

rank_adapter/adapter_model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d2715f57f10feb56f9dc22a79df65f4cad907b77f0f0c2881b110bc92db47197
+size 72400072

requirements-int4.txt

@@ -0,0 +1,2 @@
+-r requirements.txt
+bitsandbytes==0.50.0

requirements.txt

@@ -0,0 +1,7 @@
+torch>=2.5.1
+transformers==4.52.4
+peft==0.15.2
+accelerate==1.7.0
+huggingface-hub>=0.32.4
+safetensors>=0.5.3
+Pillow>=11.2.1

training_config.yaml

@@ -0,0 +1,87 @@
+project:
+  seed: 20260730
+  data_dir: data
+  output_dir: outputs
+  artifact_dir: artifacts
+
+resources:
+  card_asset_dir: asset/cards/robmikh_svg_cards
+  expected_card_count: 52
+  model_dir: model/smolvlm2-2.2b-instruct
+  card_split_counts:
+    train: 40
+    val: 6
+    test: 6
+
+geometry:
+  card_width_mm: 100.0
+  card_height_mm: 60.0
+  min_piece_count: 2
+  max_piece_count: 4
+  max_piece_edges: 5
+  min_edge_mm: 20.0
+  min_piece_area_mm2: 350.0
+  random_partition_attempts: 800
+  congruence_tolerance_mm: 0.08
+  difficult_congruence_tolerance_mm: 3.5
+  max_enumerated_candidates: 512
+  max_rank_candidates: 4
+  visual_equivalence_mae: 2.0
+
+dataset:
+  train_samples: 18000
+  val_samples: 1000
+  test_samples: 1000
+  family_weights:
+    random_unique: 0.15
+    offset_parallelogram_diagonal: 0.25
+    center_double_cut: 0.25
+    congruent_symmetric: 0.20
+    difficult_near_symmetric: 0.15
+  generation_retries: 80
+  progress_every: 100
+  binary_hard_negatives_per_scene: 1
+
+render:
+  pixels_per_mm: 6.0
+  card_width_px: 600
+  card_height_px: 360
+  board_width_px: 1280
+  board_height_px: 820
+  board_background_rgb: [226, 231, 237]
+  cell_background_rgb: [244, 246, 248]
+  label_rgb: [25, 31, 42]
+  seam_width_px_range: [1, 3]
+  piece_brightness_range: [0.94, 1.06]
+  piece_color_range: [0.96, 1.04]
+  global_brightness_range: [0.90, 1.10]
+  global_contrast_range: [0.92, 1.08]
+  noise_sigma_range: [0.0, 2.0]
+  blur_radius_range: [0.0, 0.45]
+  jpeg_quality: 94
+
+model:
+  max_length: 2048
+  max_image_longest_edge: 1280
+
+training:
+  seed: 20260730
+  learning_rate: 0.0002
+  train_batch_size_per_gpu: 1
+  eval_batch_size_per_gpu: 1
+  gradient_accumulation_steps: 4
+  lora_rank: 16
+  lora_alpha: 32
+  lora_dropout: 0.05
+  warmup_ratio: 0.05
+  logging_steps: 10
+  eval_steps: 250
+  save_steps: 250
+  save_total_limit: 2
+  dataloader_workers_per_gpu: 2
+  binary_epochs: 1
+  rank_epochs: 3
+
+evaluation:
+  max_samples: 1000
+  max_new_tokens: 4