Charles Goddard
@chargoddard
Followers
648
Following
164
Media
1
Statuses
28
Chief of Frontier Research @arcee_ai MergeKit author Github:
Joined March 2009
Don't wanna be here?
Send us removal request.
Explore trending content on Musk Viewer
Gold
• 1025064 Tweets
Modi
• 233906 Tweets
Neeraj
• 200136 Tweets
#JinxGucci
• 168282 Tweets
Cori Bush
• 139230 Tweets
中丸くん
• 94841 Tweets
WE LOVE YOU YOONGI
• 94391 Tweets
女子大生
• 52344 Tweets
#ARMYLovesSuga
• 51634 Tweets
#विनेश_फोगाट
• 49661 Tweets
アパホテル
• 42471 Tweets
KARN x DAOU
• 27915 Tweets
路上ナンパ
• 22941 Tweets
内田副総裁
• 22251 Tweets
中丸さん
• 21927 Tweets
剥離骨折
• 20658 Tweets
小池百合子知事
• 20654 Tweets
イスラエル招待
• 19904 Tweets
#earthquake
• 18671 Tweets
#Phogat_Vinesh
• 17771 Tweets
Dragoneer
• 17339 Tweets
TransisiPRABOWO JKWmulus
• 16589 Tweets
SATUarah NKRImaju
• 15938 Tweets
長崎の平和式典
• 14422 Tweets
プロ野球の始球式
• 14269 Tweets
Rossmann
• 11540 Tweets
全治2カ月
• 11083 Tweets
全員女子の戦隊
• 10767 Tweets
Always great to work with
@Teknium1
and crew. This model turned out amazing, definitely give it a try!
Today we are releasing an experimental new model in collaboration with
@chargoddard
and
@arcee_ai
, Hermes 2 Θ, our first model merge, combining Hermes 2 Pro, and Llama-3 Instruct, and then further RLHF'ed from there.
Available on HuggingFace:
This model
21
60
319
1
4
27
Hermes 2 Theta, now in 70B! Having the function calling capabilities of Hermes 2 Pro alongside the upsettingly good instruction following of Llama 3 70B Instruct is a very powerful combination - I've been having a lot of fun with this. As always a true pleasure to work with Nous.
Introducing Hermes 2 Theta 70B!
Hermes 2 Theta is smarter, more creative, and capable of more then ever before.
It takes a strong lead over Llama-3 Instruct 70B across a wide variety of benchmarks, and is a continuation of our collaboration with
@chargoddard
and
@arcee_ai
.
10
61
333
0
2
11
This is a really beautiful piece of work. WARP makes great use of properties of model merging I don't often see combined. Catastrophic forgetting mitigation, capability enhancement, KL/reward balancing, and low-bandwidth parallelization all at once? Hell yeah.
Introducing Weight Averaged Rewarded Policies (WARP), Google DeepMind's latest RLHF alignment method using the magic of model merging. By scaling alignment like pre-training was scaled, WARP learns sota Gemma LLM surpassing previous releases. A 🧵below.
6
36
231
0
4
11
Another killer paper from Sakana AI.
Can LLMs invent better ways to train LLMs?
At Sakana AI, we’re pioneering AI-driven methods to automate AI research and discovery. We’re excited to release DiscoPOP: a new SOTA preference optimization algorithm that was discovered and written by an LLM!
19
259
1K
0
0
9
Maxime quick on the draw with the first Llama 3 merge!
0
0
7
0
0
4
@MaziyarPanahi
@maximelabonne
@arcee_ai
No need to attach the whole model, a writeup of what you did and the config or a link to a huggingface page works!
0
0
2
@noguchis
どのアーキテクチャをマージを試みていますか？まだサポートされていない場合、追加を検討できますので教えてください。 また、メモリ使用量を減らすために --lazy-unpickle オプションの使用もご検討ください。 日本語は少し忘れかけていますが、お手伝いできることがあれば嬉しいです。
1
0
1
@noguchis
ありがとうございます！「LLaMAForCausalLM」は以前のtransformersのクラス名でしたが現在は変更されています。モデルのconfig.jsonのarchitecturesアレイを["LlamaForCausalLM"]にかわればmergekitは分かります。問題はすみません！未来にmergekitは「LLaMAForCausalLM」もわかるようになるはずです。
1
0
1