Zagora - LLM Fine-Tuning Platform

Email We'll notify you when your job is complete

Model

LoRA Strategy

Attention Only Standard LoRA on attention projections (q, k, v, o) Expert LoRA Targets MoE expert parameters (gate_up_proj, down_proj)

Training Configuration

Quick Test Limited data 1 epoch · 500 examples · 512 max length Full Training Full dataset training 1 epoch · all examples · 1024 max length

Dataset Link Paste a Google Drive, Dropbox, or HuggingFace link. Always include https://

Supported Dataset Formats

Chat (recommended)

{"messages": [{"role": "user", "content": "..."}, {"role": "assistant", "content": "..."}]}

Instruction

{"instruction": "Your prompt here", "response": "Expected response"}

Pre-formatted

{"text": "Complete formatted text as a single string"}

DPO (preference tuning)

{"chosen": [{"role": "user", "content": "..."}, {"role": "assistant", "content": "..."}],
 "rejected": [...], "ref_chosen_logps": -1.5, "ref_rejected_logps": -3.2}

Format is auto-detected from your JSONL. One JSON object per line.

QLoRA Fine-Tuning Beta