lfm-trainer 0.1.0

Fine-tune LFM 2.5 1.2B for coding tasks on Kaggle multi-GPU with auto-publish to Hugging Face

LFM Trainer

Fine-tune Liquid LFM 2.5 1.2B for coding tasks on Kaggle multi-GPU — with automatic checkpoint publishing to Hugging Face on errors.

Features

• 🚀 Multi-GPU training via HuggingFace Accelerate / DDP
• 🧠 LoRA / PEFT for memory-efficient fine-tuning on Kaggle T4s
• 📦 Structured dataset loading — CSV, Parquet, JSONL, or HuggingFace Hub IDs — with auto-format detection
• 🛡️ Error-resilient training — auto-publishes versioned checkpoints to Hugging Face on OOM, SIGTERM (Kaggle timeout), KeyboardInterrupt, or any exception
• 🔑 Flexible HF auth — CLI arg, HF_TOKEN env var, or Kaggle Secrets

Installation

pip install lfm-trainer

Quick Start (Kaggle Notebook)

# In a Kaggle cell:
!pip install lfm-trainer

# Train on a single dataset
!lfm-train --dataset iamtarun/python_code_instructions_18k_alpaca --hub-repo your-username/lfm-code

# Train on multiple datasets
!lfm-train \
    --dataset code_data.csv \
    --dataset more_code.parquet \
    --dataset sahil2801/CodeAlpaca-20k \
    --hub-repo your-username/lfm-code \
    --epochs 3 \
    --batch-size 2

The HF token is automatically picked up from Kaggle Secrets (key: HF_TOKEN).

CLI Reference

lfm-train --help

Flag	Default	Description
--dataset	(required)	Dataset path or Hub ID (repeatable)
--model	liquid/LFM2.5-1.2B-Base	Model to fine-tune
--hf-token	auto-detect	HuggingFace token
--hub-repo	auto	Hub repo to push to
--epochs	3	Training epochs
--batch-size	2	Per-device batch size
--lr	2e-4	Learning rate
--max-seq-length	2048	Max sequence length
--lora-r	16	LoRA rank
--lora-alpha	32	LoRA alpha
--bf16	off	Use bfloat16
--simulate-error	off	Test auto-publish

How Auto-Publish Works

The training loop is wrapped in an error handler inspired by Unsloth:

1. SIGTERM (Kaggle timeout) → saves + pushes immediately, then exits
2. CUDA OOM → clears cache, saves + pushes
3. KeyboardInterrupt → saves + pushes
4. Any Exception → saves + pushes, then re-raises

Each checkpoint is tagged with a UTC timestamp (e.g., v20260302-153000) so versions never collide.

License

MIT