al-train

Fine-tune LLMs on AL (Business Central) code using training data from al-corpus.

Overview

Item	Detail
Language	Python 3.11+
Build	hatchling via pyproject.toml
Core deps	anthropic, click, pyyaml, sacrebleu
Optional deps	unsloth, trl, transformers, torch (training)
Data source	SShadowS/al-corpus
Model target	LoRA adapter via Unsloth QLoRA

Pipeline

al-corpus pairs     →  pairs.jsonl
       ↓
al-train describe   →  described_pairs.jsonl   (Claude Haiku)
       ↓
al-train format     →  train.jsonl + eval.jsonl (ChatML)
       ↓
al-train train      →  LoRA adapter            (Unsloth QLoRA)
       ↓
al-train eval       →  report                  (5 metrics)

Setup

python -m venv .venv
source .venv/Scripts/activate  # Windows Git Bash
pip install -e ".[train]"

Quick Start

# 1. Generate pairs with al-corpus
al-corpus pairs ./my-al-project -o pairs.jsonl

# 2. Generate descriptions (test with 100 first)
al-train describe pairs.jsonl -o described.jsonl -n 100

# 3. Full corpus via batch API
al-train describe pairs.jsonl -o described.jsonl --batch
al-train describe --poll described.jsonl.batch_id -o described.jsonl

# 4. Format for training
al-train format described.jsonl -o train.jsonl --eval eval.jsonl

# 5. Train
al-train train train.jsonl --eval eval.jsonl

# 6. Evaluate
al-train eval ./output/al-coder-lora --eval-set eval.jsonl

CLI Commands

Command	Input	Output	Description
al-train describe	pairs.jsonl	described_pairs.jsonl	Generate natural-language descriptions via Claude Haiku
al-train describe --batch	pairs.jsonl	described_pairs.jsonl	Submit as an Anthropic batch job
al-train describe --poll	.batch_id file	described_pairs.jsonl	Poll and retrieve a completed batch
al-train format	described_pairs.jsonl	train.jsonl, eval.jsonl	Convert to ChatML format and split train/eval
al-train train	train.jsonl	LoRA adapter	Fine-tune with Unsloth QLoRA
al-train eval	LoRA adapter + eval.jsonl	Report	Evaluate across 5 metrics

Requirements

Requirement	Notes
Python 3.11+
ANTHROPIC_API_KEY	Required for description generation
NVIDIA GPU, 24 GB+ VRAM	Required for training (.[train] extras)
al-corpus on PATH	Required for pair generation and evaluation

Key Files

File	Purpose
pyproject.toml	Package metadata, dependencies, entry points
pairs.jsonl	Raw AL code pairs from al-corpus
described_pairs.jsonl	Pairs enriched with Claude-generated descriptions
train.jsonl / eval.jsonl	ChatML-formatted split ready for training
output/al-coder-lora/	Trained LoRA adapter output directory

---

Author: Torben Leth <sshadows@sshadows.dk> — License: MIT