xlmtec
xlmtec is a powerful, modular, and interactive command-line tool for fine-tuning Large Language Models (LLMs). Multiple Techniques: Support for LoRA, QLoRA, and Prompt Tuning. Benchmarking: Built-in evaluation with ROUGE and other metrics.
Links
README
From the repo.
xlmtec
xlmtec is a command-line toolkit for fine-tuning large language models. Describe your task in plain English, get a ready-to-run config, browse HuggingFace models, and train — all from the terminal.
Features
- AI-powered config generation — describe your task, get a YAML config from Claude, Gemini, or GPT
- Model Hub browser — search and inspect HuggingFace models without leaving the terminal
- 5 fine-tuning methods — LoRA, QLoRA, Full, Instruction, DPO
- Config validation — catch errors before training starts
- Dry-run mode — preview your training plan without loading a model
- Rich terminal UI — progress bars, panels, colour output throughout
Installation
# Core (lightweight — no ML deps)
pip install xlmtec
# With training support
pip install xlmtec[ml]
# With AI suggestions (pick your provider)
pip install xlmtec[claude] # Anthropic
pip install xlmtec[gemini] # Google
pip install xlmtec[codex] # OpenAI
pip install xlmtec[ai] # All three
# Everything
pip install xlmtec[full]
Quickstart
1. Get an AI-generated config
xlmtec ai-suggest "fine-tune a small model for customer support" --provider claude
Outputs a ready-to-run YAML config and the exact command to run.
2. Browse models on HuggingFace
xlmtec hub search "bert" --task text-classification --limit 5
xlmtec hub trending
xlmtec hub info google/bert-base-uncased
3. Validate your config
xlmtec config validate config.yaml
4. Train
# Preview without loading model
xlmtec train --config config.yaml --dry-run
# Start training
xlmtec train --config config.yaml
Commands
| Command | Description |
|---|---|
xlmtec ai-suggest "<task>" | Generate a config from plain English |
xlmtec hub search "<query>" | Search HuggingFace models |
xlmtec hub info <model-id> | Show model details |
xlmtec hub trending | Top trending models |
xlmtec config validate <file> | Validate a YAML config |
xlmtec train --config <file> | Fine-tune a model |
xlmtec train --config <file> --dry-run | Preview training plan |
xlmtec recommend | Get method recommendation for your hardware |
xlmtec evaluate | Evaluate a fine-tuned model |
xlmtec benchmark | Compare multiple runs |
xlmtec merge | Merge LoRA adapter into base model |
xlmtec upload | Upload model to HuggingFace Hub |
xlmtec --version | Show installed version |
Fine-tuning methods
| Method | VRAM | Best for |
|---|---|---|
lora | Low (4–8 GB) | Most tasks, fast convergence |
qlora | Very low (4 GB) | Large models on limited hardware |
full | High (24 GB+) | Best quality, small models |
instruction | Low (4–8 GB) | Prompt/response style tasks |
dpo | Low (4–8 GB) | Preference learning from pairs |
AI Providers
Set your API key as an environment variable, then pass --provider:
export ANTHROPIC_API_KEY=sk-ant-...
xlmtec ai-suggest "summarise legal documents" --provider claude
export GEMINI_API_KEY=...
xlmtec ai-suggest "summarise legal documents" --provider gemini
export OPENAI_API_KEY=sk-...
xlmtec ai-suggest "summarise legal documents" --provider codex
Example config
model:
name: gpt2
dataset:
source: local_file
path: data/train.jsonl
lora:
r: 16
alpha: 32
target_modules: [c_attn]
training:
output_dir: output/run1
num_epochs: 3
batch_size: 4
learning_rate: 2e-4
Development
git clone https://github.com/Abdur-azure/xlmtec.git
cd xlmtec
pip install -e ".[full,dev]"
pytest tests/ -v --ignore=tests/test_integration.py
Changelog
See CHANGELOG.md for full release history.
License
MIT
Collected info
- ★ 0 stars
- Language: HTML
- Source updated: 3/19/2026
Config for your environment
Replace {MCP_ENDPOINT_URL} with this MCP’s endpoint URL (from its repo or docs above). No API key — you connect directly.
Tool
OS
Config file: ~/.cursor/mcp.json
{
"mcpServers": {
"mcp-server": {
"url": "{MCP_ENDPOINT_URL}"
}
}
}Paste into mcpServers in the config file. Restart Cursor after saving.
If this MCP is also published on mcpchannel.ai, you can subscribe from Browse and use the gateway config there instead.