Updated context

updated_context.txt

@@ -6,31 +6,21 @@ While these large models are great for general tasks, they may need to be fine-t
 
 As a treat to our amazing readers, we fine-tuned an LLM to build a chatbot that can answer natural questions about the rich content published on our blog. This article will walk you through the fine-tuning process, from building our own dataset to optimizing inference time for the fine-tuned model!
 
+
 The model powering the chatbot was trained on the task of Conversational Question Answering, which means it can answer questions regarding a specific context (in this case, the content of this blog) or previous questions or answers from the current conversation thread. Try chatting with the bot in a conversational style, just as you would with one of Tryolabs’ tech experts!
 
 Caveats
 Remember that this is just a small demo, so the answers may not be as accurate as expected; in the Improvements section, we’ll discuss how we can enhance the model’s responses! Also, the content of this blog was not included in the training set, which means that the chatbot will give you answers about new, unseen data!
-Response time may take around 10 seconds due to the use of 🤗 Hugging Face’s free-tier space, which only has access to two CPU cores. This response time is slow for a chatbot being used in production, but it's fast for a CPU deployment and for such large inputs (since the model needs to process the entire blog to generate an answer). We'll discuss optimizing inference time in the Optimizing with ONNX section.
+Response time may take around 10 seconds due to the use of 🤗 Hugging Face’s free-tier space, which only has access to two CPU cores. This response time is slow for a chatbot being used in production, but it's fast for a CPU deployment and for such large inputs (since the model needs to process the entire blog to generate an answer).
 The Boom of Foundation Models
-Have you ever wondered how computers can understand and respond to human language? The answer lies in a new concept called ‘Foundation Models’. Popularized under this name by the Stanford University, these are machine learning models trained on immense amounts of data (over tens of terabytes and growing) that can be adapted to a wide range of downstream tasks in fields such as image, text, and audio, among others.
+Popularized under this name by the Stanford University, these are machine learning models trained on immense amounts of data (over tens of terabytes and growing) that can be adapted to a wide range of downstream tasks in fields such as image, text, and audio, among others.
 
 LLMs, a type of Foundation Model, are particularly well-established in Natural Language Processing (NLP). These models are cutting-edge artificial intelligence, capable of producing human-like language with remarkable accuracy. They can be used to generate text for various tasks such as translation, summarization, paraphrasing, question answering, and even assisted coding.
 
-These models come in different varieties and are often open-source. They can vary in size, the tasks they were pre-trained on, efficiency, and other characteristics. For example, the smallest LLMs have only a few million parameters, while the largest ones, such as the recently released Switch Transformers, have over a trillion. 
-
-Some of the most widely adopted LLMs are:
-
-BERT, developed by Google
-BLOOM, developed by BigScience Workshop
-GPT-J, the open-source version of GPT-3 developed by OpenAI
-OPT, developed by Meta
-T5, developed by Google
-A comprehensive list of open-source language models can be found on 🤗 Hugging Face’s list of supported models. Thanks to 🤗 Hugging Face and the ML community for providing this huge open-source repository!
+These models come in different varieties and are often open-source. They can vary in size, the tasks they were pre-trained on, efficiency, and other characteristics.
 
 One recent example of the impressive capabilities of Foundation Models is ChatGPT, arguably one of the most incredible advances in AI history. This model, developed by OpenAI, is a fine-tuned version of GPT-3.5 (one of the latest versions of the GPT-3 model family). ChatGPT can be used through a simple chat interface to perform various tasks, including summarization, text generation, code generation, and question-answering on virtually any topic. What sets ChatGPT apart is its ability to produce highly detailed, complete, and human-like responses, admit mistakes and adjust responses based on the user's needs.
 
-As the name suggests, Foundation Models can serve as the foundation for many applications, but they require fine-tuning and adjustments to get the most out of them. So keep reading to learn how to use them to their full potential! 
-
 Fine-tuning an LLM
 1. Building a Dataset
 There is a wide range of public datasets that you can use to fine-tune your model, but if you desire to improve the results for your specific task, you will probably need to build your own dataset. This might be a crucial task to achieve good results, albeit time-consuming.
@@ -44,7 +34,7 @@ The dataset and guidelines for building it are available in the following GitHub
 
 
 2. Choosing a Foundation Model
-Selecting the right open-source LLM to fine-tune can be a tough choice. There’s a handful of them, and while they may perform quite well for a broad range of general tasks, it can be challenging to predict how well they will perform on our specific task before fine-tuning. Think of it like buying a new pair of dancing shoes - they may look great and feel comfortable, but you won't know how well they'll perform until you hit the dance floor.
+Selecting the right open-source LLM to fine-tune can be a tough choice. There’s a handful of them, and while they may perform quite well for a broad range of general tasks, it can be challenging to predict how well they will perform on our specific task before fine-tuning.
 
 In addition to wanting the model to perform well on our specific task, we also wanted to consider other factors, such as resource efficiency and the size of the dataset required for fine-tuning. We wanted a model that could run on minimal resources and be fine-tuned with a small dataset.
 
@@ -71,7 +61,6 @@ How much computing power is available for training and inference. This can be a
 
 
 3. Fine-tuning strategy
-Once we selected the perfect Foundation Model, the next step was to prepare our data to be able to train on it. 
 
 One of the challenges we faced was that our dataset was relatively small, which could make it harder for the model to learn how to converse and answer questions about our blog posts.
 
@@ -92,6 +81,7 @@ In addition to deciding how to combine the datasets, we also needed to adapt our
 To refresh the reader’s memory, in the use-case of Conversational Question Answering, the goal of the model is to generate the answer to a question given a context and the previous questions and answers from the conversation. With this in mind, we formatted the inputs following this structure:
 
 
+
 Here, the input is a text string containing the context (i.e., the content of one of Tryolabs blog posts) followed by the last two question-and-answer pairs in the conversation and the current target question, with the target output being the answer to the target question.  We chose to add just the last two question-and-answer pairs to limit the amount of conversation history the model needs to pay attention to while still being able to generate coherent responses. Note that this is a hyper-parameter you can adjust when fine-tuning your own model. 
 
 With our data prepared and fine-tuning strategy determined, the final step was setting up our infrastructure environment and training the model.
@@ -121,8 +111,14 @@ To assess its performance, we used the F1 Score by validating how many tokens ap
 Since we had two different training steps, we also had two additional evaluation steps. The first training, on SQuAD2.0 and CoQa, resulted in a 74.29 F1 Score on the validation split after 3 epochs. The second training, on TryoCoQa, produced a 54.77 F1 Score after 166 epochs.
 
 
+
+
+
+
 More than analyzing the quantitative metrics is required to evaluate these results and conversational models in general. It is essential to consider the qualitative aspect of the model's answers, like their grammatical correctness and coherence within the conversation context. Sometimes it is preferable to have better answers (qualitatively speaking) than a better F1. So we looked at some answers from the validation set to ensure that the model was correctly generating what we were looking for. Our analysis revealed that higher F1 scores were generally associated with greater-quality answers. As a result, we selected the checkpoint with the highest F1 score to use in constructing our demonstration chatbot.
 
+If you want to play around with our fine-tuned model, you can find it on 🤗 Hugging Face with the ID tryolabs/long-t5-tglobal-base-blogpost-cqa!
+
 Faster inference with 🤗 Optimum and ONNX
 After fine-tuning our model, we wanted to make it available to our awesome readers, so we deployed it on 🤗 Hugging Face Spaces, which offers a free tier with two CPU cores for running inference on the model. However, this setup can lead to slow inference times, and processing significant inputs like ours doesn’t make it any better. And a chatbot that takes a few minutes to answer a question doesn't strike anyone as being particularly chatty, does it? So, to improve the speed of our chatbot, we turned to 🤗 Optimum and the ONNX Runtime! 
 
@@ -130,16 +126,13 @@ After fine-tuning our model, we wanted to make it available to our awesome reade
 In our previous blog post, A guide to optimizing Transformer-based models for faster inference, we used 🤗 Optimum and ONNX to achieve an x8 speed-up on inference for a Transformer model. Be sure to check it out!
 Using 🤗 Optimum’s recently released exporters feature, we were able to convert our PyTorch model to the ONNX format. This feature is handy for encoder-decoder models like the LongT5 model we trained, as it exports the three main components separately: the encoder, the decoder with the Language Modeling head, and the same decoder with pre-computed hidden states as additional inputs. According to 🤗 Optimum’s documentation, combining these three components can speed up sequential decoding, which results in faster text generation.
 
+Our fine-tuned model, exported to ONNX into these three components, is also available on 🤗 Hugging Face with the ID tryolabs/long-t5-tglobal-base-blogpost-cqa-onnx!
+
 Once our model was exported to ONNX, we used 🤗 Optimum’s integration with the ONNX Runtime to optimize our model and run inference on it by using the ORTModelForSeq2SeqLM class. This class can optimize and downcast the model using ONNX’s tools and then use ONNX Runtime to run inference with this new, faster model! You can even take it one step further and quantize the model for even shorter inference time on CPU and lower memory consumption.
 
-With these improvements, we achieved an x2 speed-up on inference time! Although the model still takes around 10 seconds to answer, this is a reasonable speed for a CPU-only deployment and processing such large inputs.
+With these improvements, we could achieve an x2 speed-up on inference time! Although the model still takes around 10 seconds to answer, this is a reasonable speed for a CPU-only deployment and processing such large inputs.
 
-Improvements
-We’re just scratching the surface of what’s possible. There are numerous potential improvements to keep working on to enhance the user's overall experience and the chatbot's performance. Here are some ideas to keep working on:
 
-Fine-tune on more public datasets for different downstream tasks. By doing so, we hope that the model learns to respond in a more human-like manner. For example, we could fine-tune on context-free QA datasets or on free-form answers so that the model not only retrieves the answers from the context but also generates original answers. We could also fine-tune on summarizing or paraphrasing tasks so that the user can also ask the chatbot to rewrite certain parts of the blog or summarize entire sections!
-Increase the size of our dataset. We built a small dataset for this demo, so it's not surprising that the model doesn't generalize well to new unseen blogs. We could potentially train a more accurate model by providing it with more training data.
-Optimize training using ONNX Runtime. We could achieve better results in the same amount of time if we could train for more epochs or fit a larger version of LongT5 in the same amount of memory. To do this, we could leverage the optimization power of ONNX Runtime for training PyTorch models, which would allow us to accelerate training speeds and optimize memory usage. Thanks to the recently released 🤗 Optimum integration with ONNX Runtime for training, this can be done in just a few lines of code!
 Takeaways
 With the ever-increasing popularity of LLMs, it can seem almost impossible to train these models without having access to millions of dollars in resources and tons of data. However, with the right skills and knowledge about Foundation Models, Deep Learning, and the Transformer architecture, we showed you that fine-tuning these huge models is possible, even with few resources and a small dataset!