dashai-frankenstein

A DashAI plugin that registers Frankenstein Transformer model classes as DashAI components, so end users can train, evaluate, predict, save, and load them from the DashAI UI.

Components registered

Schema (v1: passthrough YAML)

Each model exposes a minimal pydantic schema with a single user-facing field: frankenstein_yaml, a string containing a full Frankenstein training YAML. The Frankenstein JSON Schema is the source of truth — the YAML is validated against it (additionalProperties: false + enums) and Frankenstein's config loader (cross-component constraints) before any train/inference launches. Errors surface to the DashAI user as a readable ValueError.

Build your YAML with the Frankenstein YAML builder and paste it into the field.

Training parameters (device, batch_size, num_epochs, learning rate) are read from the YAML's training_runtime block and optimizer parameters — they are NOT separate DashAI form fields. Generation parameters (max_new_tokens, temperature, top_k) on the decoder component are kept as DashAI fields (they are inference-time, not training-time, and the Frankenstein schema has no home for them).

Note: The field is a single-line text input (DashAI does not yet support a multiline textarea for plugin schema fields). A true multiline textarea is tracked as a future upstream improvement to DashAI.

Install

pip install dashai-frankenstein            # from PyPI once published
# or, from this repo:
pip install -e ./dashai-frankenstein

DashAI discovers the plugin via the dashai.plugins entry-points group on startup — no DashAI source edits required.

Architecture

See docs/dashai-plugin-audit.md in the Frankenstein repo for the full integration design (§5 component designs, §6 phased plan, §7 Frankenstein changes). This package is the Phase 1–3 adapter layer; it consumes the Frankenstein engine API (src.engine) added in Phase 0.