# 模型平台

**Repository Path**: cqszkj/model-platform

## Basic Information

- **Project Name**: 模型平台
- **Description**: No description available
- **Primary Language**: Unknown
- **License**: Not specified
- **Default Branch**: main
- **Homepage**: None
- **GVP Project**: No

## Statistics

- **Stars**: 0
- **Forks**: 0
- **Created**: 2026-07-14
- **Last Updated**: 2026-09-25

## Categories & Tags

**Categories**: Uncategorized

**Tags**: None

## README

# Model Platform

Standalone operations platform for preparing corpora, training GPU models, tracking progress, evaluating candidates, and exporting deployment-compatible artifacts.

The first release supports:

- Student and journal Chinese/English AI-vs-human adversarial classifiers.
- The 11-family `style_analysis` attribution ensemble consumed by `aigc-gpu-detector`.
- The 11-family `attribution_laya` fine-tune: Laya multilingual base, RLCD plus soft-label
  cross-entropy over 300-character chunks, with post-training temperature calibration.
- An independent seven-family domestic attribution evaluation that never trains or registers a model.
- Resumable browser uploads, dataset/model CRUD and downloads, a MySQL-backed GPU0 queue, live job events, evaluation gates, and automatic candidate registration.

Laya training needs `transformers` 5.x, which conflicts with the worker's `transformers<5` pin, so
the worker image builds a separate `/opt/laya` venv (`--system-site-packages`, reusing the image's
torch). The Laya recipe runs there; the main environment and every other recipe are untouched.

The platform never writes into a production model directory. Successful training runs atomically register a versioned candidate archive with checksums, lineage, metrics, and its training manifest in the model library. Successful evaluation runs only retain immutable reports and do not change the model library.

## Quick start

```bash
cp .env.example .env
docker compose -f compose.yml up --build -d
```

Open `http://localhost:18090` and sign in with the administrator credentials from `.env`.

## Server deployment

The production server uses a stable self-updating entrypoint. The standard update mode builds the
new release, interrupts queued or active training jobs during the final switch, verifies all four
services, external MySQL, and GPU0, and restores the previous application images if the new release
fails health checks.

```bash
bash /devdata/disk_ssd1/docker/model-platform-script/deploy.sh --interrupt-active
```

First-time server setup and exact verification commands are documented in
[`deploy/README.md`](deploy/README.md).

Detailed deployment, data layout, model contracts, and verification commands are documented under `docs/`.
The seven-family evaluation procedure and its fixed 0820 baseline inputs are documented in
[`docs/domestic7-evaluation.md`](docs/domestic7-evaluation.md).
The Chinese `attribution_v2` upload, training, replay, development evaluation, and one-time blind
evaluation workflow is **deprecated** — its recipes can no longer create new jobs — and is kept as a
historical record in
[`docs/attribution-v2-training-guide.zh-CN.md`](docs/attribution-v2-training-guide.zh-CN.md).