henry1477/pcbslm-static-v2-unsloth-vlm-gui

---
license: other
task_categories:
- image-text-to-text
- visual-question-answering
language:
- en
pretty_name: PCBSLM static-v2 Unsloth VLM
tags:
- unsloth
- gemma-4
- multimodal
- pcb
- electronics
size_categories:
- 1K<n<10K
configs:
- config_name: default
data_files:
- split: train
path: data/vlm_train.jsonl
- split: validation
path: data/vlm_val.jsonl
- split: test
path: data/vlm_test.jsonl
---

# PCBSLM static-v2 Unsloth VLM

Portable multimodal Unsloth dataset for PCB layout/document-grounded training.

The JSONL splits use Unsloth/Gemma-style chat messages:

```json
{
"messages": [
{"role": "user", "content": [
{"type": "image", "image": "https://huggingface.co/datasets/henry1477/pcbslm-static-v2-unsloth-vlm-gui/resolve/main/assets/raw_docs/.../images/page.png"},
{"type": "text", "text": "instruction..."}
]},
{"role": "assistant", "content": [
{"type": "text", "text": "{...json answer...}"}
]}
]
}
```

## Files

- `data/vlm_train.jsonl`: 1455 multimodal examples
- `data/vlm_val.jsonl`: 100 multimodal examples
- `data/vlm_test.jsonl`: 135 multimodal examples
- `assets/raw_docs/`: source documents, rendered pages, and figure crops referenced by examples/metadata
- `assets/board_images/`: board render images referenced by examples
- `metadata_bundle.tar.gz`: document/evidence/rule metadata with repo-relative asset paths
- `asset_manifest.jsonl`: image refs with asset kind, local existence, and dimensions
- `quality_dropped_rows.jsonl`: deterministic filter drops and reasons
- `reports/quality_audit.md`: duplicate, citation, image mix, and docpack quality audit

Rows include `quality_score`, `quality_weight`, `quality_reasons`, and `mixture_bucket`. Use
`quality_weight` or `mixture_bucket` in your sampler to keep document-grounded rows near 35-45%
of training batches without duplicating examples.

## Unsloth Smoke Test

This was verified locally with `unsloth/gemma-4-E2B-it` using:

```bash
python scripts/smoke_train_unsloth_vlm.py \
--dataset-repo-id henry1477/pcbslm-static-v2-unsloth-vlm-gui \
--model-name unsloth/gemma-4-E2B-it \
--limit 4 \
--max-steps 2 \
--max-images 1 \
--max-seq-length 512 \
--resize 256
```

Image entries are HTTPS URLs to files in this dataset repo. The repository includes the referenced document and board assets.

HF repo: `henry1477/pcbslm-static-v2-unsloth-vlm-gui`

---
许可证：其他
任务类别：
- 图像-文本转文本
- 视觉问答
语言：
- 英语
友好展示名称：PCBSLM static-v2 Unsloth 视觉语言模型（Vision-Language Model, VLM）
标签：
- unsloth
- gemma-4
- 多模态
- 印刷电路板（Printed Circuit Board, PCB）
- 电子学
样本规模类别：
- 1000 < n < 10000
配置：
- 配置名称：default
数据文件：
- 拆分集：训练集
路径：data/vlm_train.jsonl
- 拆分集：验证集
路径：data/vlm_val.jsonl
- 拆分集：测试集
路径：data/vlm_test.jsonl
---

# PCBSLM static-v2 Unsloth 视觉语言模型（Vision-Language Model, VLM）

适用于印刷电路板（Printed Circuit Board, PCB）布局与文档驱动训练的可移植多模态Unsloth数据集。

本数据集采用JSONL格式拆分，使用Unsloth/Gemma风格的对话消息格式：
json
{
"messages": [
{"role": "user", "content": [
{"type": "image", "image": "https://huggingface.co/datasets/henry1477/pcbslm-static-v2-unsloth-vlm-gui/resolve/main/assets/raw_docs/.../images/page.png"},
{"type": "text", "text": "instruction..."}
]},
{"role": "assistant", "content": [
{"type": "text", "text": "{...json answer...}"}
]}
]
}

## 文件说明
- `data/vlm_train.jsonl`：包含1455条多模态样本
- `data/vlm_val.jsonl`：包含100条多模态样本
- `data/vlm_test.jsonl`：包含135条多模态样本
- `assets/raw_docs/`：存放样本与元数据所引用的源文档、渲染页面及截图裁剪区域
- `assets/board_images/`：存放样本所引用的电路板渲染图像
- `metadata_bundle.tar.gz`：包含文档、佐证材料与规则元数据，附带仓库相对路径的资产索引
- `asset_manifest.jsonl`：记录图像引用信息，包含资产类型、本地存在状态与图像尺寸
- `quality_dropped_rows.jsonl`：记录经确定性过滤剔除的样本及剔除原因
- `reports/quality_audit.md`：包含重复样本、引用关系、图像混合及文档包质量的审计报告

每条样本包含`quality_score`（质量评分）、`quality_weight`（质量权重）、`quality_reasons`（质量归因）与`mixture_bucket`（混合桶）字段。在采样器中可使用`quality_weight`或`mixture_bucket`，将文档驱动样本的占比控制在训练批次的35%-45%区间内，同时避免样本重复。

## Unsloth 冒烟测试
本数据集已通过`unsloth/gemma-4-E2B-it`在本地完成冒烟测试，测试命令如下：
bash
python scripts/smoke_train_unsloth_vlm.py
--dataset-repo-id henry1477/pcbslm-static-v2-unsloth-vlm-gui
--model-name unsloth/gemma-4-E2B-it
--limit 4
--max-steps 2
--max-images 1
--max-seq-length 512
--resize 256

图像条目均为指向本数据集仓库内文件的HTTPS链接，本仓库包含所有被引用的文档与电路板资产。

Hugging Face仓库地址：`henry1477/pcbslm-static-v2-unsloth-vlm-gui`