---
base_model: Qwen/Qwen3-4B-Instruct-2507
library_name: mlx
tags:
  - lora
  - mlx
  - trip-optimizer
  - travel
  - judge
license: apache-2.0
---

# trip-optimizer-judge

MLX LoRA adapter for **Qwen3-4B-Instruct-2507** that assesses violation candidates
in travel itinerary proposals for [trip-optimizer](https://github.com/michaelpersonal/trip-optimizer).

## Task

Given a trip profile, a proposed itinerary change, and deterministically computed
violation candidates, the model assesses each candidate as `valid`, `invalid`, or
`uncertain`, with severity (`none/minor/major/critical`) and fixability
(`none_needed/simple_revision/major_rewrite/terminal`). It also outputs any
additional violations it detects independently. The final accept/reject/revise
decision is derived programmatically by trip-optimizer from these assessments.

Input task field: `assess_violations`

## Usage with mlx-lm

```bash
pip install mlx-lm
```

```python
from mlx_lm import load, generate

model, tokenizer = load(
    "Qwen/Qwen3-4B-Instruct-2507",
    adapter_path="esc511/trip-optimizer-judge",
)
```

## Usage with trip-optimizer

```bash
# After installing trip-optimizer
trip-optimizer setup   # pulls via Ollama
```

Or point directly at this adapter by running mlx-lm with it and configuring
the endpoint in `~/.trip-optimizer/config.json`.

## Training details

- **Base model:** Qwen/Qwen3-4B-Instruct-2507
- **Method:** MLX LoRA (r=8, alpha=20, dropout=0)
- **Layers:** 32
- **Training iters:** 600
- **Max sequence length:** 4096
- **Hardware:** Apple M5 Max 128 GB
- **Checkpoint:** v8 (trip-judge-v1-mlx-r8-v8)

## Input / Output schema

Full schemas at:
- Input: [trip_judge_v1_input.schema.json](https://github.com/michaelpersonal/trip-optimizer/blob/main/finetuning/judge/schemas/trip_judge_v1_input.schema.json)
- Output: [trip_judge_v1_output.schema.json](https://github.com/michaelpersonal/trip-optimizer/blob/main/finetuning/judge/schemas/trip_judge_v1_output.schema.json)