Files
BoostAI/FineTune/README.md
2026-05-26 13:43:09 +01:00

82 lines
1.9 KiB
Markdown

# FineTune Helper
Local helper app for creating assignment-level fine-tuning records.
## Start
1. Copy `.env.example` to `.env`
2. Fill in the hosted AI endpoint, key, and model
3. Fill in the backend generator URL and teacher token
4. Run from repo root:
```bash
make fine-tune
```
Then open:
```text
http://localhost:4310
```
## What it does now
- generates a full assignment from the real backend `POST /api/questions/generate` endpoint using:
- `topic`
- `difficulty`
- `count`
- stores the assignment in the same shape the real review flow expects:
- assignment metadata
- question list
- student submission per question
- teacher review per question
- assignment summary
- recommended next step
- can ask the hosted model to draft:
- the full student submission for all questions
- the full teacher review package for all questions plus assignment summary
- shows:
- a canonical saved record preview
- a chat-style fine-tune JSON preview
- saves reviewed examples locally in your browser
- lets you load, update, and delete saved examples
- exports either:
- `dataset.jsonl`
- `train.jsonl` + `val.jsonl`
## Saved record shape
The helper now targets one saved row per assignment:
```text
assignment-review-v1
assignment
studentSubmission
teacherReview.questions[]
teacherReview.assignmentSummary
teacherReview.recommendedNextStep
```
This matches the real app's mixed-granularity review flow:
```text
one assignment review call
-> question-level labels for every question
-> one assignment-level summary
```
## Backend generator auth
Set:
- `FINE_TUNE_BACKEND_URL` to your BoostAI base URL, for example `https://boost.ai.moku.build`
- `FINE_TUNE_BACKEND_TOKEN` to a valid teacher JWT/session token value
The helper forwards that token as:
```text
Authorization: Bearer <token>
```
so it can call the protected backend generator endpoint from the separate local helper app.