FrontAgent Planner 14B LoRA MLX

This repository contains a FrontAgent Planner-only LoRA adapter trained with MLX-LM on Apple Silicon. It generates structured frontend engineering execution plans for FrontAgent. It does not directly generate business code, run tools, or make repository changes. Generated plans still require FrontAgent schema validation and runtime safety checks.

Artifact

Adapter format: MLX-LM LoRA
Adapter weights: adapters.safetensors
Recommended MLX base: mlx-community/Qwen2.5-Coder-14B-Instruct-4bit
Upstream base model: Qwen/Qwen2.5-Coder-14B-Instruct

Training Summary

Data: 500 v2 train samples and 100 v2 eval samples
MLX split: 500 train, 80 valid, 20 test
Total training: 1000 iterations (50-iter stable pilot + 950 continuation iterations)
Learning rate: 1e-5
LoRA: all layers, rank 32, alpha/scale 64
Max sequence length: 2048
Peak memory observed on local Mac: 13.249 GB
Final test loss: 0.000
Final test perplexity: 1.000

Usage with MLX-LM

python -m venv .venv
.venv/bin/python -m pip install mlx-lm
.venv/bin/python -m mlx_lm generate \
  --model mlx-community/Qwen2.5-Coder-14B-Instruct-4bit \
  --adapter-path ceilf6/frontagent-planner-14B-lora \
  --system-prompt '你是一个资深前端工程师和项目规划专家。只输出 JSON，不要输出 markdown。' \
  --prompt '任务：创建一个用户登录页面，包含邮箱和密码输入框，支持表单验证\n项目上下文：React 18 + TypeScript + Ant Design 5' \
  --max-tokens 1536 \
  --temp 0

Limitations

This adapter is MLX-LM format, not a PEFT/Transformers adapter.
It is optimized for FrontAgent planning JSON, not general coding.
Generated plans must be schema-checked before execution.
Loss metrics are from the deterministic v2 dataset split; run FrontAgent generated-plan quality evaluation before production rollout.