Libre AI Models #12
Loading…
Add table
Add a link
Reference in a new issue
No description provided.
Delete branch "%!s()"
Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?
Determine which are the SOTA libre models at various scales.
See:
https://spacecruft.org/deepcrayon/parrot-models
Check which of these models used in Ollama have good licenses:
codellama:34b-code
codellama:34b-code-q8_0
codellama:34b-instruct-q8_0
codellama:34b-python
codellama:34b-python-q8_0
everythinglm:13b-16k-q8_0
falcon:40b
falcon:40b-instruct
falcon:40b-instruct-fp16
falcon:40b-text
falcon:40b-text-fp16
falcon:180b
falcon:180b-chat-q4_0
falcon:180b-text-q4_0
falcon:180b-text-q8_0
llama2:70b
mistral:7b-instruct-q8_0
mistral:7b-text-q8_0
nexusraven:13b
nexusraven:13b-q8_0
orca-mini:latest
phind-codellama:34b-python
phind-codellama:34b-python-q8_0 9e8121c0c614 36
phind-codellama:34b-v2
phind-codellama:34b-v2-q2_K
phind-codellama:34b-v2-q4_0
phind-codellama:34b-v2-q4_1
phind-codellama:34b-v2-q4_K_S
phind-codellama:34b-v2-q6_K
phind-codellama:34b-v2-q8_0
sqlcoder:15b
sqlcoder:15b-fp16
sqlcoder:15b-q8_0
starcoder:15b-base-q8_0
vicuna:33b
vicuna:33b-q8_0
wizardcoder:7b-python
wizardcoder:13b-python
wizardcoder:34b-python
wizardcoder:34b-python-q2_K
wizardcoder:34b-python-q6_K
wizardcoder:34b-python-q8_0
wizard-math:70b
wizard-math:70b-q5_K_M
Llama's license isn't libre.
Llama non-free license includes these:
NexusRaven
Codellama
Phind-CodeLlama
WizardCoder
Vicuna
WizardLM
CodeLlama
BigCode OpenRAIL-M and OpenRAIL++ are non-free licenses, includes these:
CodeUp-Llama
StarCoder
WizardCoder-15B-V1.0
Other proprietary licenses:
Claude
GPT
PaLM
Qwen
StableCode Research License is non-free:
StableCode
Possibly libre, under apache 2 license:
Falcon
Mistral
Pythia
Possibly libre, under CC-BY-SA:
SQLCoder
Stable LM
Open-Assistant StableLM
Phind 34b v2 works well with llama.cpp and parrot with continuedev extension. It has fake apache license that makes it non-libre.
Deepseek coder is new, afaict. 33b under MIT, trained on code.
Meh, deepseek says MIT, but has a fake open source license for the model.
This may be actual "open source":
https://github.com/EleutherAI
https://www.eleuther.ai
I don't see a coding model, but it has many others, and tools, which perhaps could make one.
EleutherAI trains on "The Pile"
https://pile.eleuther.ai/
That full dataset is ~850GB, apache license, available via torrent, etc.
"Older" model (2022), but appears to actually be libre model (apache 2.0 license, without BS):
https://beta.the-eye.eu/public/AI/models/GPT-NeoX-20B/
Eleuther AI claims "open source", and some parts may be, but it also uses proprietary models, such as llama license, eg.
https://huggingface.co/EleutherAI/llemma_7b
So they aren't all "open source" either, need to double check anything licensed from them.
Coqui license is non-libre.
chatglm3-6b license:
"You will not use the Software for any act that may undermine China's national security and national unity"
Another "open source" license that isn't open source:
Yi Series Models Community License Agreement
Maybe
https://huggingface.co/TigerResearch
In the last few months there have been many new models released and along with that, more license proliferation.
The following models have been tested and used with Parrot, on an AMD CPU server with six AMD Radeon RX 7900 XTX (tinyrocs). Most have also been tested on Nvidia A16s and/or Nvidia A6000s. Using llama.cpp server with GGUF files, either from upstream, TheBloke, or custom built.
Nous Research Llama 3 70B Instruct
Qwen CodeQwen1.5 7B Chat
Qwen1.5 110B Chat
CodeLlama 70B Instruct
Mistral 7B v0.1
Phind CodeLlama 34B v2
WizardCoder Python 34B v1.0
Deepseek Coder 33B Instruct
Mixtral 8x7B Instruct
OpenCodeInterpreter DS 33B
StarCoder2 3B
Falcon 40B Instruct (on Nvidia?)
CodeQwen is by far the fastest, and is the highest rated on the huggingface code leaderboard for Python. It is from Alibaba and under a unique Alibaba license (e.g. corporate fake open source license).
(Pseudo-) open source models recommended by continuedev:
Chat (e.g. code generation):
llama3 70b, 8b
deepseek-coder 33b, 6.7b
Autocomplete:
deepseek-coder 6.7b, 1.3b
StarCoder2 7b, 3b
Embeddings:
transformers.js
nomic-embed-text