point-alpaca  by pointnetwork

Fine-tuned LLaMA weights, recreated from Stanford Alpaca

Created 2 years ago
404 stars

Top 71.9% on SourcePulse

GitHubView on GitHub
1 Expert Loves This Project
Project Summary

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.

Health Check
Last Commit

2 years ago

Responsiveness

1 day

Pull Requests (30d)
0
Issues (30d)
0
Star History
0 stars in the last 30 days

Explore Similar Projects

Starred by Jiayi Pan Jiayi Pan(Author of SWE-Gym; MTS at xAI), Chip Huyen Chip Huyen(Author of "AI Engineering", "Designing Machine Learning Systems"), and
12 more.

EasyLM by young-geng

0.0%
2k
LLM training/finetuning framework in JAX/Flax
Created 2 years ago
Updated 1 year ago
Starred by Chip Huyen Chip Huyen(Author of "AI Engineering", "Designing Machine Learning Systems"), Shizhe Diao Shizhe Diao(Author of LMFlow; Research Scientist at NVIDIA), and
17 more.

open_llama by openlm-research

0.1%
8k
Open-source reproduction of LLaMA models
Created 2 years ago
Updated 2 years ago
Starred by Andrej Karpathy Andrej Karpathy(Founder of Eureka Labs; Formerly at Tesla, OpenAI; Author of CS 231n), George Hotz George Hotz(Author of tinygrad; Founder of the tiny corp, comma.ai), and
20 more.

TinyLlama by jzhang38

0.1%
9k
Tiny pretraining project for a 1.1B Llama model
Created 2 years ago
Updated 1 year ago
Starred by Junyang Lin Junyang Lin(Core Maintainer at Alibaba Qwen), Vincent Weisser Vincent Weisser(Cofounder of Prime Intellect), and
25 more.

alpaca-lora by tloen

0.0%
19k
LoRA fine-tuning for LLaMA
Created 2 years ago
Updated 1 year ago
Starred by Andrej Karpathy Andrej Karpathy(Founder of Eureka Labs; Formerly at Tesla, OpenAI; Author of CS 231n), John Yang John Yang(Coauthor of SWE-bench, SWE-agent), and
28 more.

stanford_alpaca by tatsu-lab

0.1%
30k
Instruction-following LLaMA model training and data generation
Created 2 years ago
Updated 1 year ago
Starred by Roy Frostig Roy Frostig(Coauthor of JAX; Research Scientist at Google DeepMind), Zhiqiang Xie Zhiqiang Xie(Coauthor of SGLang), and
40 more.

llama by meta-llama

0.1%
59k
Inference code for Llama 2 models (deprecated)
Created 2 years ago
Updated 7 months ago
Feedback? Help us improve.