Please welcome K2-65B🏔️, the most performant fully-open LLM released to date.
As a blueprint for open-source AGI, we release all model checkpoints, code, logs, and data.
About K2:
🧠65 billion parameters
🪟Fully transparent & reproducible
🔓Apache 2.0
📈Outperforms Llama 2 70B
🚀 1/7 We are thrilled to launch LLM360 — pushing the frontier of open-source & transparent LLMs!
Starting with Amber (7B) & CrystalCoder (7B), we are releasing brand new pre-trained LLMs with all training code, data, and up to 360 model checkpoints.
🔗
1/3 We are releasing CrystalChat 🔮 — a top-scoring 7B chat model, fully open source!
As always, CrystalChat is released under Apache 2.0, along with all training data, checkpoints, and implementation details.
Grab the model here:
🔍 2/7 By releasing all data, code, & checkpoints, LLM360 makes it easy to reproduce results and build off our models for research and industry purposes.
All models are released under Apache 2.0 license.
Learn more
Blog:
Paper:
1/3 We’re adding two new models to LLM360! Presenting AmberChat 💬 & AmberSafe 🦺.
AmberChat is instruction tuned on Amber (7B) using
@WizardLM_AI
&
@ShareGPT
data.
AmberSafe is DPO FT’d AmberChat using PKU-SafeRLHF data.
AC:
AS:
🌟 3/7 Amber, a 7B English LLM, is pre-trained on 1.2T tokens. See Amber’s emergent capabilities over 360 checkpoints & dive into 6.8TB of data.
Model:
Metrics:
Data:
Code:
🔮 4/7 CrystalCoder, a 7B code & text model, combines the best of StarCoder & Llama. Trained w/ 1.4T tokens on
@CerebrasSystems
Condor Galaxy 1.
Model:
Metrics:
Data:
Code:
🤝 5/7 We are excited to continue contributing to the community through open releases. Join us through direct collaboration or by telling us how we can help.
LLM360 is proudly brought to you by
@PetuumInc
,
@MBZUAI
, and
@CerebrasSystems
.
Huge congratulations to the
@allen_ai
team!
The OLMo series is a substantial contribution to the OSS LLM community with:
- open training datasets
- 500 checkpoints
- analysis and evaluations
Couldn’t be happier to see more projects with similar goals (cc
@AiEleuther
)
OLMo is here! And it’s 100% open.
It’s a state-of-the-art LLM and we are releasing it with all pre-training data and code. Let’s get to work on understanding the science behind LLMs. Learn more about the framework and how to access it here:
We evaluated K2-65B across 22 standard benchmarks to assess its broad knowledge on topics such as coding, medicine, and math, in addition to Open LLM Leaderboard metrics.
🔗Check out the model here:
K2 was generously sponsored by
@MBZUAI
and
@PetuumInc
.
🎉 Congratulations to an awesome fully open source model, by the m-a-p team!
Paper: 📎
Includes great info on:
-Data Curation
-Infra details
-Intermediate checkpoints
-Scaling law
LLM360 is happy to work with this thriving community on open source AI.
🚀 Excited to announce that the tech report of MAP-Neo (): a fully open-source and transparent bilingual LLM suite with superior performance to bridge the gap with closed-source models, is now available:
🔧MAP-Neo's workflow
We also released a fine-tuned chat model, K2-Chat.
K2-Chat outperforms Llama 2 70B-Chat in medicine and math metric groups, and outperforms Llama 3 70B-Instruct on coding tasks.
We open-source a variety of components in three suites 📕
1. *LLM360 Research Suite*: artifacts such as model ckpts, code, and data.
2. *LLM360 Pretraining Suite*: tutorials to reproduce and build on our models.
3 . *LLM360 Developer Suite*: tutorials on fine-tuning, running
❄️Congrats to
@SnowflakeDB
for openly releasing Arctic!❄️
Arctic is available to all with an Apache 2.0 license!
Great to see LLM360 member
@AurickQ
and the whole Snowflake AI Research’s team's amazing contribution to the open-source LLM community!
2/3 Choosing models without knowing the training data is becoming riskier and riskier (e.g.
@nytimes
vs
@openai
).
CrystalChat makes all that information available — with more to come!
Check out how CrystalChat compares to SOTA models and performs on the major benchmarks below.
🔥Congrats to
@MaitrixOrg
for releasing Pandora, a World Model that can predict and simulate the world’s states in visual space, controllable by language 🎮
Excited to see the MaitrixOrg is indeed building something like the Matrix, great work by
@szxiangjn
,
@guangyi_l
,
@YiGu025
🔥Introducing Pandora 🌏 🪐
a World Model that generates videos of world states with real-time language control 🎥🕹️
Simulate the world across domains in an _interactive_ way!
check out more
Join our AMA in the Mozilla AI discord this Thursday (5/30) at 3pm EST.
We will answer questions about our models, how we trained them, and anything in between.
Thanks to Mozilla for the invite!
Link in the thread below.
@amasad
Check out — we are releasing fully-transparent OSS LLMs with:
- full code/implementation
- open data
- large sets of checkpoints
- open weights
and more, for a large set of models (7B-65B) and specialities (english, code, chat, etc).
🔥Congrats to
@MaitrixOrg
for their new library: LLM Reasoners v1.0🔥
Great to see
@Ber18791531
and LLM360 members
@YiGu025
and
@ZhitingHu
’s continued contributions to open-source LLMs!
Check them out:
🔎The LLM360 Research Suite was developed to aid academic and industry researchers further the ability and understanding of LLMs
📈We release detailed artifacts so everyone has access to the same material, as if they are model trainers
Research Suite artifacts highlighted below
Introducing Meta Llama 3: the most capable openly available LLM to date.
Today we’re releasing 8B & 70B models that deliver on new capabilities such as improved reasoning and set a new state-of-the-art for models of their sizes.
Today's release includes the first two Llama 3
3/3 CrystalChat is on
@huggingface
It is fine-tuned from CrystalCoder-7B, originally trained on
@CerebrasSystems
Condor Galaxy 1.
Stay tuned as we peel back the onion on CrystalChat and show you everything under the hood.
As always, drop us a line on ✍️
@karpathy
🙏 for the shoutout! It’s amazing to hear such a positive response from everyone in the community on transparent and open-source LLM research!
3/3 Both models are available on
@huggingface
.
Special shout out to
@TheBlokeAI
for quantization!
We'd love to hear how LLM360 can do more to foster a transparent, trustworthy, and collaborative ecosystem.
Drop us a line in the feedback form on ✍️
2/3 Comparing AmberChat and AmberSafe side-by-side shows additional work is needed to ensure all LLMs are safe for human usage.
Transparent safety alignment will continue the progress of open-source LLMs.
Thanks to
@AnthropicAI
and others for opening their alignment data!
@WizardLM_AI
Thank you very much! We also use your evol-instruction sets, they work well with our models! We will be posting some results about the fine tuning too.
Loss spikes are a poorly understood phenomena.
We encountered two malignant loss spikes resulting in significant performance degradation while training K2-65B.
We saved the spike checkpoints for further evaluation.
Malignant spikes:
We release full details to produce K2-65B, Crystal-7B, and Amber-7B.
Artifacts include:
- 600 total intermediate checkpoints
- full data sequence and dataset
- training and data prep code
- evaluation code and results
Training dynamics: