xlmtec 3.15.0

Production-grade LLM fine tuning framework with CLI

xlmtec

Production-grade LLM fine tuning, distillation, and pruning from the command line.

---

What it does

xlmtec is a modular Python framework for fine-tuning, distilling, and pruning large language models. It wraps HuggingFace Transformers + PEFT in a clean CLI, a validated config system, a composable trainer stack, an interactive TUI, and a full test suite — all CPU-runnable for unit tests.

---

Install

git clone https://github.com/Abdur-azure/xlmtec.git
cd xlmtec
pip install -e .

---

5-minute quickstart

# 1. Generate sample training data (no network required)
python examples/generate_sample_data.py

# 2. Not sure which method to use? Ask
xlmtec recommend gpt2 --output my_config.yaml

# 3. Train with the generated config
xlmtec train --config my_config.yaml

# 4. Or use a ready-made config
xlmtec train --config examples/configs/lora_gpt2.yaml

# 5. Launch the interactive TUI
xlmtec tui

---

CLI commands

Command	What it does
xlmtec train	Fine-tune using a YAML config or inline flags (LoRA / QLoRA / Full / Instruction / DPO / Distillation)
xlmtec evaluate	Score a saved checkpoint (ROUGE, BLEU, Perplexity)
xlmtec benchmark	Before/after comparison: base vs fine-tuned
xlmtec merge	Merge LoRA adapter into base model → standalone model
xlmtec upload	Push adapter or merged model to HuggingFace Hub
xlmtec recommend	Inspect model size + VRAM, output optimal YAML config
xlmtec prune	Structured pruning — zero lowest-magnitude attention heads
xlmtec wanda	WANDA unstructured pruning — zero weights by |W|×activation score
xlmtec tui	Interactive Textual TUI — all commands via a terminal UI

---

Training methods

Method	Flag	Notes
LoRA	--method lora	Default. Adapter-based, memory-efficient
QLoRA	--method qlora	4-bit quantised LoRA — large models on limited VRAM
Full Fine-Tuning	--method full_finetuning	All parameters — small models only
Instruction Tuning	--method instruction_tuning	Alpaca-style {instruction, input, response} data
DPO	--method dpo	Direct Preference Optimization — requires pip install trl
Response Distillation	--method vanilla_distillation	Student mimics teacher logits (KL + CE loss)
Feature Distillation	--method feature_distillation	Student mimics teacher hidden states (MSE + KL + CE)

---

Pruning commands

# Structured pruning — zero lowest-magnitude attention heads
xlmtec prune ./outputs/gpt2_lora \
    --output ./outputs/gpt2_pruned \
    --sparsity 0.3 \
    --method heads

# WANDA unstructured pruning — weight × activation scoring, zero-shot
xlmtec wanda ./outputs/gpt2_lora \
    --output ./outputs/gpt2_wanda \
    --sparsity 0.5 \
    --dataset ./data/sample.jsonl

---

Example configs

Config	Method	Model	Data
lora_gpt2.yaml	LoRA	GPT-2	data/sample.jsonl
qlora_llama.yaml	QLoRA	LLaMA-3.2-1B	HF Hub (needs token)
instruction_tuning.yaml	Instruction	GPT-2	data/instructions.jsonl
full_finetuning.yaml	Full	GPT-2	data/sample.jsonl
dpo.yaml	DPO	GPT-2	data/dpo_sample.jsonl
response_distillation.yaml	Response Distillation	GPT-2 (student) ← GPT-2-medium	data/sample.jsonl
feature_distillation.yaml	Feature Distillation	GPT-2 (student) ← GPT-2-medium	data/sample.jsonl
structured_pruning.yaml	Structured Pruning	GPT-2	—
wanda.yaml	WANDA Pruning	GPT-2	data/sample.jsonl (calibration)

---

Python API

from xlmtec.core.config import ConfigBuilder
from xlmtec.core.types import TrainingMethod, DatasetSource
from xlmtec.models.loader import load_model_and_tokenizer
from xlmtec.data import prepare_dataset
from xlmtec.trainers import TrainerFactory

config = (
    ConfigBuilder()
    .with_model("gpt2")
    .with_dataset("./data/sample.jsonl", source=DatasetSource.LOCAL_FILE)
    .with_tokenization(max_length=256)
    .with_training(TrainingMethod.LORA, "./output", num_epochs=3)
    .with_lora(r=8, lora_alpha=16)
    .build()
)

model, tokenizer = load_model_and_tokenizer(config.model.to_config())
dataset = prepare_dataset(config.dataset.to_config(), config.tokenization.to_config(), tokenizer)
result = TrainerFactory.train(
    model, tokenizer, dataset,
    config.training.to_config(),
    config.lora.to_config(),
)
print(f"Done. Loss: {result.train_loss:.4f}  →  {result.output_dir}")

---

Docs

• Usage Guide — all 9 commands with examples
• Configuration Reference — YAML config fields for all methods
• API Reference — Python API for all trainers and pruners
• TUI Guide — interactive terminal interface
• Architecture — module design
• Contributing — how to add trainers or commands

---

Tests

# Unit tests (no GPU needed)
pytest tests/ -v --ignore=tests/test_integration.py

# Integration tests (CPU ok, ~30s — downloads GPT-2 once)
pytest tests/test_integration.py -v -s

# Full suite
pytest tests/ -v

---

Project status

Aspect	Status
Version	3.13.0
Tests	200+ unit + integration, all green
CI	pytest on Python 3.10 / 3.11 / 3.12
Platform	Windows / macOS / Linux
License	MIT