Zain Hasan
@zainhasan6.bsky.social
I build & teach AI stuff | AI/ML @ http://Together.ai (ex Weaviate) | ℕΨ @UofT Engineering | Data Scientist | Lecturer 🇨🇦🇵🇰
Views not mine!!
Views not mine!!
Pinned
Cooking up something nice for y'all!👨🍳
Can't wait to share more soon...
Can't wait to share more soon...
June 15, 2025 at 3:38 PM
Cooking up something nice for y'all!👨🍳
Can't wait to share more soon...
Can't wait to share more soon...
this is so cool - shows the real time molecular-scale interaction of hydrogen + oxygen atoms merging to form nano-sized bubbles of water🫧
December 1, 2024 at 4:46 AM
this is so cool - shows the real time molecular-scale interaction of hydrogen + oxygen atoms merging to form nano-sized bubbles of water🫧
Improving Summarization for long documents using long-context fine-tuning!
New Cookbook: Long Document Summarization + Evaluation.
We fine-tune Llama 3.1 8B to improve summarization of documents 32k tokens long and show outperformance over 70B models!
New Cookbook: Long Document Summarization + Evaluation.
We fine-tune Llama 3.1 8B to improve summarization of documents 32k tokens long and show outperformance over 70B models!
November 28, 2024 at 5:12 AM
Improving Summarization for long documents using long-context fine-tuning!
New Cookbook: Long Document Summarization + Evaluation.
We fine-tune Llama 3.1 8B to improve summarization of documents 32k tokens long and show outperformance over 70B models!
New Cookbook: Long Document Summarization + Evaluation.
We fine-tune Llama 3.1 8B to improve summarization of documents 32k tokens long and show outperformance over 70B models!
New technical deep dive into multi-turn conversation fine-tuning!
We fine-tune Llama 3.1 8B, with instruction loss masking, on the conversational CoQA dataset and show a 2x improvement in exact match score for outputs!
We fine-tune Llama 3.1 8B, with instruction loss masking, on the conversational CoQA dataset and show a 2x improvement in exact match score for outputs!
November 27, 2024 at 3:33 AM
New technical deep dive into multi-turn conversation fine-tuning!
We fine-tune Llama 3.1 8B, with instruction loss masking, on the conversational CoQA dataset and show a 2x improvement in exact match score for outputs!
We fine-tune Llama 3.1 8B, with instruction loss masking, on the conversational CoQA dataset and show a 2x improvement in exact match score for outputs!
How do you teach an LLM to carry long form conversations and not get confused by all the details?
To learn how we can improve fine-tuning over long form conversational data I fine-tuned a bunch of models on the CoQA dataset and 2x'd performance!
Full code notebook below🔽
To learn how we can improve fine-tuning over long form conversational data I fine-tuned a bunch of models on the CoQA dataset and 2x'd performance!
Full code notebook below🔽
November 27, 2024 at 3:31 AM
How do you teach an LLM to carry long form conversations and not get confused by all the details?
To learn how we can improve fine-tuning over long form conversational data I fine-tuned a bunch of models on the CoQA dataset and 2x'd performance!
Full code notebook below🔽
To learn how we can improve fine-tuning over long form conversational data I fine-tuned a bunch of models on the CoQA dataset and 2x'd performance!
Full code notebook below🔽
Definition of generative agents.
an architecture that extends a LLM to store a complete record of the agent's experiences using natural language, synthesize those memories over time into higher-level reflections, and retrieve them dynamically to plan behavior.
an architecture that extends a LLM to store a complete record of the agent's experiences using natural language, synthesize those memories over time into higher-level reflections, and retrieve them dynamically to plan behavior.
November 25, 2024 at 5:31 AM
Definition of generative agents.
an architecture that extends a LLM to store a complete record of the agent's experiences using natural language, synthesize those memories over time into higher-level reflections, and retrieve them dynamically to plan behavior.
an architecture that extends a LLM to store a complete record of the agent's experiences using natural language, synthesize those memories over time into higher-level reflections, and retrieve them dynamically to plan behavior.
Main use cases of LLMs
November 24, 2024 at 10:56 PM
Main use cases of LLMs
Great talk from Prof. Potts on thinking of LLMs as part of systems rather than just the LLM alone.
He covers:
> Prompt design systems - DSPy
> Sampling systems - top_p, dynamic temp etc.
> Tool use - databases, search, functions
> Evaluating system vs. LLMs
He covers:
> Prompt design systems - DSPy
> Sampling systems - top_p, dynamic temp etc.
> Tool use - databases, search, functions
> Evaluating system vs. LLMs
November 24, 2024 at 10:55 PM
Great talk from Prof. Potts on thinking of LLMs as part of systems rather than just the LLM alone.
He covers:
> Prompt design systems - DSPy
> Sampling systems - top_p, dynamic temp etc.
> Tool use - databases, search, functions
> Evaluating system vs. LLMs
He covers:
> Prompt design systems - DSPy
> Sampling systems - top_p, dynamic temp etc.
> Tool use - databases, search, functions
> Evaluating system vs. LLMs
77% of enterprise AI usage are using models that are small models, less than 13b parameters.
November 24, 2024 at 10:25 PM
77% of enterprise AI usage are using models that are small models, less than 13b parameters.
Reposted by Zain Hasan
Comprehensive talk on what it takes to create LLMs!
Basically a degree in LLM creation, everything is study-able and transparent!
OLMo - open language model
OLMoE -mixture of experts
Dolma - full pre-training dataset
Molmo/PixMo - VLM
Tulu 3 - post training and datasets
Basically a degree in LLM creation, everything is study-able and transparent!
OLMo - open language model
OLMoE -mixture of experts
Dolma - full pre-training dataset
Molmo/PixMo - VLM
Tulu 3 - post training and datasets
November 22, 2024 at 5:57 AM
Comprehensive talk on what it takes to create LLMs!
Basically a degree in LLM creation, everything is study-able and transparent!
OLMo - open language model
OLMoE -mixture of experts
Dolma - full pre-training dataset
Molmo/PixMo - VLM
Tulu 3 - post training and datasets
Basically a degree in LLM creation, everything is study-able and transparent!
OLMo - open language model
OLMoE -mixture of experts
Dolma - full pre-training dataset
Molmo/PixMo - VLM
Tulu 3 - post training and datasets
This paper is Westworld vibes for real.
Generative Agent Simulations of 1,000 People - arxiv.org/abs/2411.10109
The generative agents replicate participants' responses 85% as accurately as participants replicate their own answers two weeks later.
Generative Agent Simulations of 1,000 People - arxiv.org/abs/2411.10109
The generative agents replicate participants' responses 85% as accurately as participants replicate their own answers two weeks later.
a woman is holding a book and asking if she is real
ALT: a woman is holding a book and asking if she is real
media.tenor.com
November 22, 2024 at 6:17 AM
This paper is Westworld vibes for real.
Generative Agent Simulations of 1,000 People - arxiv.org/abs/2411.10109
The generative agents replicate participants' responses 85% as accurately as participants replicate their own answers two weeks later.
Generative Agent Simulations of 1,000 People - arxiv.org/abs/2411.10109
The generative agents replicate participants' responses 85% as accurately as participants replicate their own answers two weeks later.
Comprehensive talk on what it takes to create LLMs!
Basically a degree in LLM creation, everything is study-able and transparent!
OLMo - open language model
OLMoE -mixture of experts
Dolma - full pre-training dataset
Molmo/PixMo - VLM
Tulu 3 - post training and datasets
Basically a degree in LLM creation, everything is study-able and transparent!
OLMo - open language model
OLMoE -mixture of experts
Dolma - full pre-training dataset
Molmo/PixMo - VLM
Tulu 3 - post training and datasets
November 22, 2024 at 5:57 AM
Comprehensive talk on what it takes to create LLMs!
Basically a degree in LLM creation, everything is study-able and transparent!
OLMo - open language model
OLMoE -mixture of experts
Dolma - full pre-training dataset
Molmo/PixMo - VLM
Tulu 3 - post training and datasets
Basically a degree in LLM creation, everything is study-able and transparent!
OLMo - open language model
OLMoE -mixture of experts
Dolma - full pre-training dataset
Molmo/PixMo - VLM
Tulu 3 - post training and datasets
Manhattan Project-like program to race to AGI.
November 20, 2024 at 4:55 AM
Manhattan Project-like program to race to AGI.
Qwen2.5-Turbo is cracked!
📚 1M context length with 100% accuracy and 4x cheaper then 4o -mini
📚 1M context length with 100% accuracy and 4x cheaper then 4o -mini
November 20, 2024 at 4:36 AM
Qwen2.5-Turbo is cracked!
📚 1M context length with 100% accuracy and 4x cheaper then 4o -mini
📚 1M context length with 100% accuracy and 4x cheaper then 4o -mini
Kind of insane to think of what the best LLMs can currently do and then realize this is the worst they will ever be.
"what's already happened is much more important than anything else that's going to be done and then it's just going to be a long ways in applying it." - Thiel
"what's already happened is much more important than anything else that's going to be done and then it's just going to be a long ways in applying it." - Thiel
November 20, 2024 at 4:23 AM
Kind of insane to think of what the best LLMs can currently do and then realize this is the worst they will ever be.
"what's already happened is much more important than anything else that's going to be done and then it's just going to be a long ways in applying it." - Thiel
"what's already happened is much more important than anything else that's going to be done and then it's just going to be a long ways in applying it." - Thiel
LLaVA-o1, a novel VLM designed to conduct autonomous multistage reasoning. Unlike chain-of-thought prompting, LLaVA-o1 independently engages in sequential stages of summarization, visual interpretation, logical reasoning, and conclusion generation.
November 20, 2024 at 4:22 AM
LLaVA-o1, a novel VLM designed to conduct autonomous multistage reasoning. Unlike chain-of-thought prompting, LLaVA-o1 independently engages in sequential stages of summarization, visual interpretation, logical reasoning, and conclusion generation.
"tokens" a 30 year old human would have "trained" on amounts to ~31,728 T
Counts visual, tactile, auditory, olfactory, taste data.
Interesting hypothesis.
Counts visual, tactile, auditory, olfactory, taste data.
Interesting hypothesis.
November 20, 2024 at 4:15 AM
"tokens" a 30 year old human would have "trained" on amounts to ~31,728 T
Counts visual, tactile, auditory, olfactory, taste data.
Interesting hypothesis.
Counts visual, tactile, auditory, olfactory, taste data.
Interesting hypothesis.