Jaigouk Kim (@jaigouk) header
Jaigouk Kim (@jaigouk) profile picture

a geek

Joined January 2008
Hermes 3 performs strongly against Llama-3.1 Instruct Models, but with a focus on aligning the model to you, instead of a company or external policy - meaning less censorship and more steerability - with additional capabilities like agentic XML, scratchpads, roleplaying prowess, and more. Step level reasoning and planning, internal monologues, improved RAG, and even LLM as a judge capabilities were also targeted. Below are benchmark comparisons between Hermes 3 and Llama-3.1 Instruct and a sample of utilizing the agentic XML tags:
9
21
10
260
21,516
Introducing `llama-405b-to-8b` ✍️ Get the quality of Llama 3.1 405B, at a fraction of the cost and latency. Give one example of your task, and 405B will teach 8B (~30x cheaper!!) how to do the task perfectly. And it's open-source: github.com/mshumer/gpt-promp…
29
272
37
2,134
278,599
You don't need a H100 to run Llama-3-405b. 2 MacBooks and 1 Mac Studio will do the job, with @exolabs_ to aggregate the memory/compute. I'm ready for you, Llama-3-405b.
38
106
34
956
104,364
Jaigouk Kim retweeted
Released Hermes 2 Pro on Llama-3 8B today! Get it here: huggingface.co/NousResearch/… or the GGUF here: huggingface.co/NousResearch/…
Announcing Hermes 2 Pro on Llama-3 8B! Nous Research's first Llama-3 based model is now available on HuggingFace. Hermes Pro comes with Function Calling and Structured Output capabilities, and the Llama-3 version now uses dedicated tokens for tool call parsing tags, to make handling streaming function calls easier. This model surpasses Llama-3 8B Instruct on AGIEval, GPT4All Suite, TruthfulQA, and BigBench. Check it out here: huggingface.co/NousResearch/… GGUF quantized versions available here: huggingface.co/NousResearch/… The Hermes Pro models were built from the efforts and collaboration of @teknium, @intrstllrninja, @theemozilla, @karan4d, and @huemin_art; thanks as well to @ByteNano and @winglian for customizing axolotl to support our unique prompt formatting.
41
80
8
634
85,226
Jaigouk Kim retweeted
🚨 Breaking news: Google just introduced ScreenAI, and it's wild. This is going to transform the future of UX forever Here's everything you need to stay ahead of the curve: 🧵 👇
105
647
72
3,884
1,498,765
Jaigouk Kim retweeted
A crew writing crewAI docs, preparing the next version release 😎👉👉 and it does get more meta than this in some of out other crews
7
5
2
114
6,380
The way we do normal lora won't work bitnet, bcs we have to maintain property when lora weight are merged they can still be quantized into good ternary, we need littel reformulation in forward pass, so BitLora
4
16
1
113
10,336
We are releasing our first step in validating and independently confirming the claims of the Bitnet paper, a 1B model trained on the first 60B tokens of the Dolma dataset. Comparisons made on the @wandb charts below are between the Bitnet implementation and a full FP16 run (all hyperparameters equivalent). Model: huggingface.co/NousResearch/… Weights & Biases: api.wandb.ai/links/emozilla/…
24
151
31
894
227,518
276
24,494
778
222,400
5,676,499
Jaigouk Kim retweeted
Open-sourcing "MindGraph", a proof of concept starter kit for building and querying against an ever-expanding knowledge graph with natural language! - multiple input types - simple interface - basic search/RAG against knowledge graph - easily extendable with integrations - basic database swapping GitHub in 🧵
44
270
26
1,813
243,407
you can now run function calling and json mode with @ollama thanks to @AdrienBrault 🔥
I have created and pushed @ollama models for Hermes 2 Pro 7B! ollama.com/adrienbrault/nous…
1
6
67
7,382
Introducing the latest version of our Hermes series of models, Hermes 2 Pro 7B. This latest version improves several capabilities, using an updated and cleaned version of the Hermes 2 dataset, and is now trained on a diverse and rich set of function calling and JSON mode samples, thus opening the door to much more reliable and robust agents. Download the model here: huggingface.co/NousResearch/… GGUF versions available here: huggingface.co/NousResearch/… This work was months in the making, and was built by the efforts and collaboration of @teknium, @intrstllrninja , @theemozilla, @karan4d, @huemin_art, and @FireworksAI_HQ; thanks as well to @ByteNano and @winglian for customizing axolotl to support our unique prompt formatting A special thanks to @Latitudesh for sponsoring the compute to make this model.
24
104
20
589
138,121
Jaigouk Kim retweeted
(Q)LoRA in MLX LM is also faster and more memory efficient thanks to: - compilation - better data packing - gradient checkpointing pip install -U mlx-lm Fine-tuning 4-bit Mistral 7B on an 8GB (!) M1 is actually quite doable:
3
37
4
259
20,317
Today, with @Tim_Dettmers, @huggingface, & @mobius_labs, we're releasing FSDP/QLoRA, a new project that lets you efficiently train very large (70b) models on a home computer with consumer gaming GPUs. 1/🧵 answer.ai/posts/2024-03-06-f…
78
623
101
3,406
445,276
Today we are announcing the release of a new model, Genstruct 7B. Genstruct 7B is an instruction-generation model, designed to create valid instructions given a raw text corpus. This enables the creation of new, partially synthetic instruction finetuning datasets from any raw-text corpus. Download the model + notebook on our HuggingFace page, here: huggingface.co/NousResearch/… Inspired by the paper, Ada-Instruct, we took this approach further by grounding the generations in user-provided context passages. Further, the model is trained to generate questions involving complex scenarios that require detailed reasoning, allowing for models trained on the generated data to reason step-by-step. This project was led by @Euclaise_ at Nous Research.
12
78
14
535
128,034
Why wait for someone else to productize your life? With open models and readily available wearable hardware, you can accelerate your cognition right now, That's why @EthanSutin and I are launching a fully open source project to build an ambient smart assistant system that listens to and observes your life, endowing you with super human memory that's fully in your control. Please reach out to either of us if you're interested in pushing the boundaries of always-on AI. We're going to launch with off-the-shelf inexpensive hardware you can turn into a necklace, pin, or any wearable form factor you want. It also works with amazing hardware you may already have, such as the Apple Watch. Expect more next week, including a GitHub repo. We will be building in public.
92
197
41
2,065
315,391
Using GPT-4 to rerank search results 🥸 I want to share two nuggets from my initial experiment: 1. Adding an example of what the output should look like helps the model follow the instructions. 2. Manually tweaking the role description to improve the ranking quality wasn't as straightforward. The visual shows how to structure a prompt for listwise reranking using OpenAI's API. This includes the role description, output format, and task example. I tested 4 prompts using 16 @weaviate_io FAQs. This lets you measure recall at 1 because each question has a ground truth answer. The best performing prompt followed the output format 100% of the time (🥳), but only got 75% accuracy across the 16 FAQs. Excited to see how I can improve this with DSPy next! @lateinteraction Notebook: github.com/weaviate/recipes/…
12
63
5
359
59,559
OpenAI's GPT Store is here. But most GPTs will be useless (I know because I went through hundreds). If you want to make a useful GPT, rank high, and make money, you need to use actions. So, here's an advanced tutorial on how to add actions to your custom GPTs:
103
287
24
1,876
544,662