Chaos Genius’ Post

At Databricks we just released DBRX, the best-in-class Open Large Language Model as of now. The model has a Mixture of Experts (MoE) architecture, that's roughly twice the size (132B) but half the cost (36B) of Llama2-70B. Since only 36B expert parameters are used at inference, it's close to twice the speed (tokens/seconds) of Llama2-70B. This remarkable achievement has been brought to life using a single Data Intelligence Platform (Databricks of course), for data ingestion, preparation, training, evaluation, serving, and governance. Check out the details https://lnkd.in/dhhwwAT9

Snowflake has unveiled an innovative open-source large language model Here are the key highlights: - Snowflake has launched Arctic, an open-source large language model (LLM) designed to enable users to build enterprise-grade AI models and applications. - Arctic uses a unique Dense-MoE Hybrid transformer architecture. It combines a 10B dense transformer model with a residual 128×3.66B MoE MLP resulting in 480B total and 17B active parameters  - Snowflake claims Arctic is the most open, enterprise-grade LLM available, and also one of the most powerful. - The model is available under an Apache 2.0 license, with its weights, code, data recipes and research insights open-sourced.(Thats best thing) - Snowflake says Arctic was built in less than three months and at roughly one-eighth of the training costs of similar LLMs. - Arctic (17B active parameters) can have up to 4x less memory reads than Code-Llama 70B, and up to 2.5x less than Mixtral 8x22B (44B active parameters), leading to faster inference performance. - They also released embedding model but did not mentioned in article. - One downside I feel is that the model's size is too large, requiring 963.07 GB of storage. Article : https://lnkd.in/d-EtSHhg try Now : https://lnkd.in/dsaCj6RV HuggingFace : https://lnkd.in/dATUwS-3

❗️❗️BIG #DATABRICKS NEWS❗️❗️ This morning we dropped news of our new open-source LLM! Whats the TLDR? - Better price/performance than industry leading LLMs. - Built on Databricks. Everything you will need, we can support. - Never lose ownership of your IP. - Train on your data, your custom LLM #DBRX is a new general-purpose LLM trained from scratch using the Databricks Data Intelligence Platform. While it has 132B total parameters, with its fine-grained MoE architecture, DBRX only uses 36B at any given time. It’s great for enterprises that want to efficiently build and train LLMs on their own data. Learn more about how the Databricks Mosaic Research team built DBRX & benchmarked its performance.

Excited to see the launch, but I'm late to see dbrx-instruct, the LLM trained from the scratch by Databricks. DBRX is A transformer based decoder only model that uses next token prediction feature to predict the next token in the sequence, and it uses fine grained Mixture of experts Architecture. Some key features includes 🔖132B parameters where 36B are active on any kind of input. 🔖12T tokens on text and code 🔖16 experts and chooses 4. Use dbrx-instruct for any kind use case linke text generation, code generation etc where the outcomes are interesting to see. Here is the link to the model : https://lnkd.in/gPkTsqYF I'm excited to try it out too 😜😁

Snowflake Arctic, a large language model (LLM), utilizes a Dense-MoE Hybrid transformer architecture to enhance training efficiency and reduce costs. It employed a three-phase training curriculum focused on enterprise metrics like SQL generation and coding, allowing it to perform competitively with models trained on much higher budgets. Arctic’s architecture combines a 10B dense transformer with a 128x3.66B MoE MLP, enabling efficient training and inference by leveraging system designs that minimize communication overhead among the model’s experts. Open-sourced under the Apache 2.0 license, Arctic provides access to its model weights and training techniques, aiming to support the development of custom models within the AI community https://lnkd.in/gunTtEzM

Snowflake Launches Arctic, An Open ‘Mixture-Of-Experts’ LLM To Take On DBRX, Llama 3 Snowflake announced the launch of Arctic, a large language model (LLM) optimized for complex enterprise workloads such as SQL generation, code generation and instruction following. Touted as the “most open enterprise-grade LLM” in the market, Arctic taps a unique mixture of expert (MoE) architecture to top benchmarks for enterprise tasks while being efficient at the same time. It also delivers competitive performance across standard benchmarks, nearly matching open models from Databricks, Meta and Mistral at tasks revolving around world knowledge, common sense, reasoning and math capabilities. https://lnkd.in/ghAZ5jNn

Build and train LLMs on your own data with an open LLM created by Databricks. #DBRX is a new general-purpose LLM trained from scratch using the Databricks Data Intelligence Platform. While it has 132B total parameters, with its fine-grained MoE architecture, DBRX only uses 36B at any given time. Learn more about how the Databricks Mosaic Research team built #DBRX & benchmarked its performance.

What is better than using world class LLMS in datbricks? A world class LLM from databricks for all your needs. #DBRX is a new general-purpose LLM trained from scratch using the Databricks Data Intelligence Platform. While it has 132B total parameters, with its fine-grained MoE architecture, DBRX only uses 36B at any given time. It’s great for enterprises that want to efficiently build and train LLMs on their own data. Learn more about how the Databricks Mosaic Research team built #DBRX & benchmarked its performance.

One of those announcements that makes me feel so happy to work at Databricks... #DBRX is a new general-purpose LLM trained from scratch using the Databricks Data Intelligence Platform. While it has 132B total parameters, with its fine-grained MoE architecture, DBRX only uses 36B at any given time. It’s great for enterprises that want to efficiently build and train LLMs on their own data. Learn more about how the Databricks Mosaic Research team built #DBRX & benchmarked its performance.

Understanding Evidently-AI “Evidently helps evaluate, test, and monitor data and ML-powered systems.” It does this through 3 core components: 1️⃣ Reports.  ✔ We can view as inline in jupyter notebook cells ✔ Can also export to JSON, HTML, Dictionary, DataFrames 2️⃣ Test suites ✔ A lot of presets for the commonly used generic ones ✔ Returns a Pass / Fail  3️⃣ Dashboards ✔ Individual report cards as panels or combine individual tests ✔ Use the Cloud or Self-host ✔ Impute into database via scripts to export to external solutions like Prometheus, Grafane, etc  If you’ve used the ydata_profiling library (renamed from pandas-profiling) for pandas dataframes before, the context is pretty similar. In module-05 Monitoring of #mlopszoomcamp, we integrate 3 services into a Docker Container: ✔ Postgres database to store the metrics data ✔ Adminer for UI to our postgres database ✔ Grafana for external dashboarding examples The baseline example notebook had us use 2 months of data. First month's dataset is split to train and validate our model, and the val_data is then used as our reference. The second month's dataset is then used to simulate current data to be compared against the refence and generate some metrics for our report. Which we then save as a dictionary so that we can extract the pertinent metrics for our Dashboard. These extracted metrics can also be routed to a #postgress database that our #grafana Dashboards rely on. There’s still so much to learn about #evidentlyai, more study is required while I continue on to the next module. #mlopszoomcamp #mlops #zoomcamp #DataTalksClub image credits: from module 6.1 of Open-source ML observability course by Evidently AI