ModelStation — SLM Fine-tuning Platform (Simple Overview)

Short summary

• ModelStation turns user requirements into a compact, high-quality SLM (small tail model) by:
  1. Capturing a user query / intent,
  2. Distilling that intent against a large dataset to produce a focused custom dataset,
  3. Fine-tuning a base SLM on that distilled dataset,
  4. Producing a small, deployable model tailored to the customer's needs.

Architecture & components

• Web frontend: user interactions, login, model creation flows. See web/hooks/use-user.tsx.
• API: authentication and model management endpoints:
  • Auth: [/auth/login, /auth/logout, /auth/me] implemented in api/app/controllers/auth/routes.py.
  • Models: Create/list/delete model entries in api/app/controllers/models/routes.py.
• Semantic split / dataset generator: queries vector store (Weaviate) and builds dataset slices for a given user intent. See DatasetGenerator and scripts/semantic_split/main.py.
• Seed & embedding: dataset ingestion and embedding creation are in scripts/seed/main.py and config in scripts/seed/src/config.py.
• Fine-tuning trainer: actual fine-tune pipeline that turns distilled data into a small SLM is in scripts/finetuning_image/src/train_app/trainer.py.

High-level flow (user-friendly)

1. User submits a request to create a custom model via the web UI.
   • Frontend code uses the user context and calls the API. See web/hooks/use-user.tsx.
2. API creates a model record (status: "pending") and enqueues a training job.
   • Model creation entrypoint: api/app/controllers/models/routes.py.
3. Semantic distillation
   • The semantic-split service queries Weaviate for examples matching the user's intent (DatasetGenerator), filters/structures rows, then optionally pushes a curated dataset to Hugging Face. See DatasetGenerator and scripts/semantic_split/main.py.
4. Embedding & seeding (if needed)
   • If you need to augment or seed data, the seed scripts embed and insert samples into the vector store. See scripts/seed/main.py and scripts/seed/src/config.py.
5. Fine-tuning
   • The trainer consumes the distilled dataset, converts to the required format, and runs fine-tuning to produce a small, highly-tailored SLM. See scripts/finetuning_image/src/train_app/trainer.py.
6. Deployment & use
   • The resulting model (small tail model) is saved and can be used for inference in any project/team. The platform maintains model metadata in the API database and exposes endpoints to list/delete models. See api/app/controllers/models/routes.py.

Why this approach?

• Distillation: selecting a focused subset of the big dataset aligns training examples with the customer's policy, objective, and distribution — producing a compact dataset that yields an efficient "tail" model.
• Small tail model: fine-tuning a small SLM on the distilled dataset produces a lightweight model that preserves desired behavior while being cheaper and faster to serve.

Developer quickstart (local)

• The repo includes Docker Compose files to run services:
  • development: compose.dev.yaml
  • production/single-file examples: compose.yaml, compose.prod.yaml
• Typical steps:
  1. Configure environment variables (do not commit secrets).
  2. Start services with Docker Compose (use the dev compose for local testing).
  3. Use the frontend to create a model or call API endpoints directly.
  4. Monitor logs for the semantic split, embedding, and trainer services.

Key endpoints & symbols

• Auth routes: api/app/controllers/auth/routes.py (login, logout, /me).
• Model management: api/app/controllers/models/routes.py (create/list/delete).
• Semantic dataset generation: DatasetGenerator and scripts/semantic_split/main.py.
• Trainer: scripts/finetuning_image/src/train_app/trainer.py.
• Seed & embedding: scripts/seed/main.py and scripts/seed/src/config.py.

Notes & next steps

• Secrets: keep API keys and tokens out of version control (use environment variables).
• Training orchestration: currently the API enqueues training entries; integrate a job-worker or pod orchestration to run the trainer and report progress back to the API.
• Metrics & QA: add evaluation suites and logging during training to validate model behavior against the customer's success criteria.

License & contribution

• See LICENSE.md for license details and README.md for contribution pointers.