This repository provides the fully fine-tuned weights for the Point-Alpaca language model, a recreation of Stanford's Alpaca experiment. It targets researchers and developers looking to leverage a powerful instruction-following LLM, offering a significant improvement over the original Alpaca through extensive fine-tuning on a synthetic dataset.

How It Works

Point-Alpaca is a full fine-tune of the LLaMA model, trained for three epochs on an 8x A100 80GB setup. This extensive training process reduced the loss from approximately 2 to 0.5, aiming for enhanced instruction-following capabilities. The project distributes the fine-tuned weights as XOR-encrypted diffs to circumvent LLaMA's licensing restrictions, requiring users to possess the original LLaMA weights for decryption.

Quick Start & Requirements

• Install: pip3 install -r requirements.txt
• Run: python3 chat.py
• Prerequisites: Original LLaMA weights (7B version at original/7B/consolidated.00.pth), Python 3, wget or equivalent for downloading diffs.
• Hardware: 16 GB VRAM (unquantized), 8 GB VRAM (8-bit quantized). Confirmed to run on a single RTX 3090 unquantized.
• Demo: https://alpaca.point.space
• Announcement: https://twitter.com/PointNetwork/status/1637178814210908160

Highlighted Details

• Full fine-tune for 3 epochs, achieving significantly lower loss than original Alpaca.
• Weights are distributed as XOR-encrypted diffs to comply with LLaMA licensing.
• Offers a live demo for immediate testing.

Maintenance & Community

• Community support available via Telegram chat: https://t.me/pointnetworkchat
• Future model releases (e.g., 13B) are contingent on community support.

Licensing & Compatibility

• The project itself appears to be unencumbered by specific licensing terms in the README. However, the use of the fine-tuned weights is inherently tied to the licensing of the base LLaMA model, which is not permissive for commercial use.

Limitations & Caveats

The "encryption" is a simple XOR, not intended for security. Users must legally obtain and possess the original LLaMA weights to reconstruct the fine-tuned model. The README mentions 13B models are planned, but availability is uncertain.