Apppublic

jellewas/gdpr-lora-v2train

sourceHugging Faceupdated 6mo agoView on Hugging Face

0likes

App README

LoRA Training Pipeline

Train a LoRA adapter (CPT + SFT) on a HuggingFace dataset using Unsloth and push the merged model to a HuggingFace repository. Runs on GPU-enabled HuggingFace Spaces.

Setup

Required Secrets

Set these as Space secrets (Settings > Repository secrets):

Secret	Description
`HF_TOKEN`	HuggingFace token with write access to your model repo
`DATA_REPO`	HF dataset repo containing training data (e.g. `MataHariResearch/gdpr-training-data`)
`MODEL_REPO`	HF model repo to push results to (e.g. `MataHariResearch/gdpr-lora`)

Optional Secrets

Secret	Description
`WANDB_API_KEY`	Weights & Biases API key for experiment tracking

Training Hyperparameters (Environment Variables)

All have sensible defaults. Override via Space secrets or env vars.

Variable	Default	Description
`BASE_MODEL`	`Qwen/Qwen3.5-4B`	Base model to fine-tune
`CPT_RANK`	`32`	LoRA rank for continued pre-training
`CPT_ALPHA`	`64`	LoRA alpha for CPT
`CPT_LR`	`2e-4`	Learning rate for CPT
`CPT_EPOCHS`	`3`	Number of CPT epochs
`SFT_RANK`	`16`	LoRA rank for supervised fine-tuning
`SFT_ALPHA`	`32`	LoRA alpha for SFT
`SFT_LR`	`1e-4`	Learning rate for SFT
`SFT_EPOCHS`	`3`	Number of SFT epochs
`MAX_SEQ_LENGTH`	`4096`	Maximum sequence length
`BATCH_SIZE`	`4`	Per-device training batch size

Training Data Format

Your DATA_REPO dataset should contain these JSONL files:

—`corpus.jsonl` (or cpt.jsonl / pretrain.jsonl) — For continued pre-training. Each line: {"text": "..."}. If absent, CPT is skipped.
—`train.jsonl` — SFT training data. Each line: {"text": "..."} (chat-formatted).
—`val.jsonl` (or validation.jsonl) — SFT validation data. Same format as train.

Workflow

1.Create a HuggingFace Space with Docker SDK and a GPU runtime (T4 or better).
2.Set the required secrets listed above.
3.The Space builds and launches the Gradio UI.
4.Click Start Training to begin the pipeline.
5.Monitor progress in the log panel — it auto-refreshes every 5 seconds.
6.When complete, the merged model is available at your MODEL_REPO.

Pipeline Steps

1.Download training data from DATA_REPO
2.Run CPT (continued pre-training) if corpus file is present
3.Run SFT (supervised fine-tuning) with CPT adapter merged in
4.Merge all adapters into the base model
5.Upload merged model to MODEL_REPO
6.Optionally convert to GGUF and upload (requires llama.cpp in the image)