Skip to content

Latest commit

 

History

History
79 lines (64 loc) · 7.52 KB

llm_models.md

File metadata and controls

79 lines (64 loc) · 7.52 KB

LLM: Non-Exhaustive List of Models

Image Description

Proprietary LLMs

Name Chatbot Model
Creator		 Parameters
(billions)		 Release Date
GPT-3 ChatGPT OpenAI,
Microsoft		 175 (LLM) May 2020
(Chatbot) Oct. 2022
GPT-3.5 Turbo ChatGPT Premium OpenAI,
Microsoft		 < 175 April 2023
GPT-4 ChatGPT Premium
Bing AI		 OpenAI
Microsoft		 ??? April 2023
LaMDA Bard Google 137 (LLM) Jan. 2022
(Chatbot) Fev. 2023
Gopher N/A Deepmind
Google		 280 N/A
Chinchilla N/A Deepmind
Google		 70 N/A
PaLM 1 N/A Google 540 N/A
PaLM 2 Bard Google 340* May 2023
Claude Poe Anthropic ??? March 2023

*Source: CNBC

Consumer LLMs, finetuned variants and alignment datasets

LLMs

It is important to distinguish the open sourced LLM from the variant that was finetuned (RLHF) on a specific dataset of instructions (RLHF-Supervised Fine-Tuning) and/or a dataset of LLM output preferences (RLHF-Proximal Policy Optimization).

Name Model
Creator		 Finetuned Variants RLHF Variants Parameters
(billions)		 Release Date Licence
GPT-NeoX EleutherAI Dolly
OASST		 None 3
20		 Feb. 2022 Open source
Flan-T5-XL
Flan-T5-XXL		 Google Fastchat-T5 None 3
11		 Oct. 2022 Open source
LLaMA 🔥 Meta Alpaca
Vicuna
Koala
OASST
OASST-h2O
Guanaco 🔥
Wizard
WizardVicuna
WizardMega
Manticore
StableVicuna
VicUnlocked
Dromedary
Samantha
Supercot
NousHermes
Airoboros
Tulu
FinLLaMA
Baichuan		 OASST
StableVicuna		 7
13
30
65		 Feb. 2023 Non-commercial
Pythia EleutherAI Dolly
OASST		 None 3
7
12		 April 2023 Open source
StableLM Stability AI OASST None 3
7
15 (TBD)
30 (TBD)
65 (TBD)
175 (TBD)		 April 2023 Open source
MPT MosaicML Storywriter None 7 May 2023 Open source
Starcoder
(mainly for code
generation)		 HuggingFace GPTeacher
Wizard 🔥		 None 15 May 2023 OpenRAIL
(open source
with use-case
restrictions)
Falcon Technology Innovation
Institute (UAE)		 Falcon-Instruct
Samantha
Wizard		 None 7
40		 May 2023 Open source
OpenLLaMA Berkeley AI Research
 Ongoing None 3
7
13
30 (TBD)
65 (TBD)		 June 2023 Open source

Finetuned Variants of Consumer LLMs

I will mostly be checking out the finetuned GPTQ and GGML LLMs from HuggingFace user TheBloke and other official providers (e.g. LMSYS, Databricks, OpenAssistant). Models to explore:

  • LLaMA
  • Falcon
  • Starcoder
  • OpenLLaMA

Finetuning datasets

  • Alpaca Finetuning Datasets
    • Stanford Alpaca SelfInstruct 52K
      • Non-commercial
      • For less than 600$ and 3 hours of finetuning on 8 80GB A100s, researchers from Stanford University generated a dataset of 52K instructions with OpenAI's "Text-davicini-003" (GPT-3) following the SelfInstruct (Wang et al., 2022), then finetuned on those instructions.
    • Microsoft Alpaca GPT-4 52K
      • Non-commercial
      • Follows Stanford's Alpaca methodology but used here GPT-4 for generating the 52k instructions (Peng et al., 2023)
  • Databricks Dolly
    • Databricks Dolly 15K
      • Open source
      • Dataset of over 15K instructions generated by Databricks' employees following the InstructGPT methodology. The dataset contains instructions covering "brainstorming, classification, closed QA, generation, information extraction, open QA, and summarization".
  • ShareGPT
    • ShareGPT
      • Non-commercial
      • Users upload their prompts/replies from ChatGPT (any version), which can then be used for cheaply finetuning LLMs without relying on human annotators. The legality of using ShareGPT is rather murky but it seems that simply using it for research purposes, back office applications (not exposing to clients) or even commercial use might not be restricted as long as the LLM finetuned on ShareGPT doesn't compete with OpenAI's ChatGPT (not unlike the controversy where Google was accused of finetuning their chatbot Bard on GPT outputs).
    • ShareGPT Unrestricted 35K
      • Non-commercial
  • WizardLM Evol-Instruct
    • WizardLM EvolInstruct 70K
      • Non-commercial
      • 70k instructions that followed WizardLM's EvolInstruct methodology. It starts with the normal Alpaca approach for generating a dataset of instructions but then the authors ask GPT-4 to rewrite the initial instructions into more complex ones. The idea is that with more complex and especially diverse instructions, this will improve the performance of LLMs with relatively low parameter count (7B, 13B), which seems to be the case since WizardLM 7B is as of May 2023 seen as the top 7B open source LLM.
    • WizardLM EvolInstruct Unrestricted 50K
      • Non-commercial
      • The original EvolInstruct 70k filtered of OpenAI's alignment contraints ("I'm sorry...", "As an AI language model, I cannot..."), reduced down to the 50K complex instructions (following WizardLM's EvolInstruct methodology). The LLaMA 7B trained on this unrestricted dataset produces better performances than the restricted vanilla LLM, which might indicate that there is an "alignment tax" (excessive moral alignment decreases model capacity to respond to user queries).

RLHF datasets

[TBD]