|
import gradio as gr |
|
|
|
gr.Markdown(""" |
|
# Big Science Bloom is a 176B Parameter Large Language ML Model. |
|
https://www.youtube.com/watch?v=wA8rjKueB3Q |
|
https://www.youtube.com/watch?v=2MBJOuVq380&t=241s |
|
# Big Science Papers and Code - Exciting AI Developments! 🤖💻🔬 |
|
https://paperswithcode.com/paper/bloom-a-176b-parameter-open-access |
|
""") |
|
|
|
api = gr.Interface.load("models/bigscience/bloom") |
|
|
|
def complete_with_gpt(text): |
|
|
|
|
|
|
|
return text[:-100] + api(text[-100:]) |
|
|
|
|
|
with gr.Blocks() as demo: |
|
with gr.Row(): |
|
textbox = gr.Textbox(placeholder="Type here and press enter...", lines=14) |
|
with gr.Column(): |
|
btn = gr.Button("Generate") |
|
|
|
btn.click(complete_with_gpt, textbox, textbox) |
|
|
|
with gr.Row(): |
|
gr.Markdown(""" |
|
|
|
# Example on how to prompt. |
|
|
|
Create a pattern sequence of text. In this example I use language names then click generate to add each line after adding another heading for a language. |
|
|
|
English: Hi my name is Aaron. I am a computer scientist and senior principal engineer. |
|
Japanese: 私はアランです。コンピューター科学者とプログラ |
|
English: Hi my name is Aaron. I am a computer scientist and senior principal engineer. |
|
Chinese: 你好，我叫Aaron。我是一个计算机科学家和高级首席工程师。 |
|
English: Hi my name is Aaron. I am a computer scientist and senior principal engineer. |
|
Spanish: Hola, me llamo Aaron. Soy un cientifico de la computacion y un ingeniero principal |
|
English: Hi my name is Aaron. I am a computer scientist and senior principal engineer. |
|
Sanskrit: नमस्ते, मेरा नाम है Aaron. मैं एक कंप्यूटर वैज्ञानिक और वरिष्ठ प्रमुख इंजीनियर हूँ। |
|
French: Bonjour, je m'appelle Aaron. Je suis un scientifique en informatique et un ingénieur senior. |
|
|
|
|
|
## Language Models 🗣️ |
|
|
|
🏆 Bloom sets new record for most performant and efficient AI model in science! 🌸 |
|
|
|
### Comparison of Large Language Models |
|
|
|
| Model Name | Model Size (in Parameters) | |
|
| ----------------- | -------------------------- | |
|
| BigScience-tr11-176B | 176 billion | |
|
| GPT-3 | 175 billion | |
|
| OpenAI's DALL-E 2.0 | 500 million | |
|
| NVIDIA's Megatron | 8.3 billion | |
|
| Transformer-XL | 250 million | |
|
| XLNet | 210 million | |
|
|
|
## ChatGPT Datasets 📚 |
|
|
|
- WebText |
|
- Common Crawl |
|
- BooksCorpus |
|
- English Wikipedia |
|
- Toronto Books Corpus |
|
- OpenWebText |
|
|
|
## ChatGPT Datasets - Details 📚 |
|
|
|
- **WebText:** A dataset of web pages crawled from domains on the Alexa top 5,000 list. This dataset was used to pretrain GPT-2. |
|
- [WebText: A Large-Scale Unsupervised Text Corpus by Radford et al.](https://paperswithcode.com/dataset/webtext) |
|
|
|
- **Common Crawl:** A dataset of web pages from a variety of domains, which is updated regularly. This dataset was used to pretrain GPT-3. |
|
- [Language Models are Few-Shot Learners](https://paperswithcode.com/dataset/common-crawl) by Brown et al. |
|
|
|
- **BooksCorpus:** A dataset of over 11,000 books from a variety of genres. |
|
- [Scalable Methods for 8 Billion Token Language Modeling](https://paperswithcode.com/dataset/bookcorpus) by Zhu et al. |
|
|
|
- **English Wikipedia:** A dump of the English-language Wikipedia as of 2018, with articles from 2001-2017. |
|
- [Improving Language Understanding by Generative Pre-Training](https://huggingface.co/spaces/awacke1/WikipediaUltimateAISearch?logs=build) Space for Wikipedia Search |
|
|
|
- **Toronto Books Corpus:** A dataset of over 7,000 books from a variety of genres, collected by the University of Toronto. |
|
- [Massively Multilingual Sentence Embeddings for Zero-Shot Cross-Lingual Transfer and Beyond](https://paperswithcode.com/dataset/bookcorpus) by Schwenk and Douze. |
|
|
|
- **OpenWebText:** A dataset of web pages that were filtered to remove content that was likely to be low-quality or spammy. This dataset was used to pretrain GPT-3. |
|
- [Language Models are Few-Shot Learners](https://paperswithcode.com/dataset/openwebtext) by Brown et al. |
|
|
|
|
|
## Big Science Model 🚀 |
|
|
|
- 📜 Papers: |
|
1. BLOOM: A 176B-Parameter Open-Access Multilingual Language Model [Paper](https://arxiv.org/abs/2211.05100) |
|
2. Megatron-LM: Training Multi-Billion Parameter Language Models Using Model Parallelism [Paper](https://arxiv.org/abs/1909.08053) |
|
3. 8-bit Optimizers via Block-wise Quantization [Paper](https://arxiv.org/abs/2110.02861) |
|
4. Train Short, Test Long: Attention with Linear Biases Enables Input Length Extrapolation [Paper](https://arxiv.org/abs/2108.12409) |
|
5. [Other papers related to Big Science](https://huggingface.co/models?other=doi:10.57967/hf/0003) |
|
6. [217 other models optimized for use with Bloom](https://huggingface.co/models?other=bloom) |
|
|
|
- 📚 Datasets: |
|
|
|
**Datasets:** |
|
|
|
1. - **Universal Dependencies:** A collection of annotated corpora for natural language processing in a range of languages, with a focus on dependency parsing. |
|
- [Universal Dependencies official website.](https://universaldependencies.org/) |
|
2. - **WMT 2014:** The fourth edition of the Workshop on Statistical Machine Translation, featuring shared tasks on translating between English and various other languages. |
|
- [WMT14 website.](http://www.statmt.org/wmt14/) |
|
3. - **The Pile:** An English language corpus of diverse text, sourced from various places on the internet. |
|
- [The Pile official website.](https://pile.eleuther.ai/) |
|
4. - **HumanEval:** A dataset of English sentences, annotated with human judgments on a range of linguistic qualities. |
|
- [HumanEval: An Evaluation Benchmark for Language Understanding](https://github.com/google-research-datasets/humaneval) by Gabriel Ilharco, Daniel Loureiro, Pedro Rodriguez, and Afonso Mendes. |
|
5. - **FLORES-101:** A dataset of parallel sentences in 101 languages, designed for multilingual machine translation. |
|
- [FLORES-101: A Massively Multilingual Parallel Corpus for Language Understanding](https://flores101.opennmt.net/) by Aman Madaan, Shruti Rijhwani, Raghav Gupta, and Mitesh M. Khapra. |
|
6. - **CrowS-Pairs:** A dataset of sentence pairs, designed for evaluating the plausibility of generated text. |
|
- [CrowS-Pairs: A Challenge Dataset for Plausible Plausibility Judgments](https://github.com/stanford-cogsci/crows-pairs) by Andrea Madotto, Zhaojiang Lin, Chien-Sheng Wu, Pascale Fung, and Caiming Xiong. |
|
7. - **WikiLingua:** A dataset of parallel sentences in 75 languages, sourced from Wikipedia. |
|
- [WikiLingua: A New Benchmark Dataset for Cross-Lingual Wikification](https://arxiv.org/abs/2105.08031) by Jiarui Yao, Yanqiao Zhu, Ruihan Bao, Guosheng Lin, Lidong Bing, and Bei Shi. |
|
8. - **MTEB:** A dataset of English sentences, annotated with their entailment relationships with respect to other sentences. |
|
- [Multi-Task Evaluation Benchmark for Natural Language Inference](https://github.com/google-research-datasets/mteb) by Michał Lukasik, Marcin Junczys-Dowmunt, and Houda Bouamor. |
|
9. - **xP3:** A dataset of English sentences, annotated with their paraphrase relationships with respect to other sentences. |
|
- [xP3: A Large-Scale Evaluation Benchmark for Paraphrase Identification in Context](https://github.com/nyu-dl/xp3) by Aniket Didolkar, James Mayfield, Markus Saers, and Jason Baldridge. |
|
10. - **DiaBLa:** A dataset of English dialogue, annotated with dialogue acts. |
|
- [A Large-Scale Corpus for Conversation Disentanglement](https://github.com/HLTCHKUST/DiaBLA) by Samuel Broscheit, António Branco, and André F. T. Martins. |
|
|
|
|
|
- 📚 Dataset Papers with Code |
|
|
|
1. [Universal Dependencies](https://paperswithcode.com/dataset/universal-dependencies) |
|
2. [WMT 2014](https://paperswithcode.com/dataset/wmt-2014) |
|
3. [The Pile](https://paperswithcode.com/dataset/the-pile) |
|
4. [HumanEval](https://paperswithcode.com/dataset/humaneval) |
|
5. [FLORES-101](https://paperswithcode.com/dataset/flores-101) |
|
6. [CrowS-Pairs](https://paperswithcode.com/dataset/crows-pairs) |
|
7. [WikiLingua](https://paperswithcode.com/dataset/wikilingua) |
|
8. [MTEB](https://paperswithcode.com/dataset/mteb) |
|
9. [xP3](https://paperswithcode.com/dataset/xp3) |
|
10. [DiaBLa](https://paperswithcode.com/dataset/diabla) |
|
|
|
# Deep RL ML Strategy 🧠 |
|
|
|
The AI strategies are: |
|
- Language Model Preparation using Human Augmented with Supervised Fine Tuning 🤖 |
|
- Reward Model Training with Prompts Dataset Multi-Model Generate Data to Rank 🎁 |
|
- Fine Tuning with Reinforcement Reward and Distance Distribution Regret Score 🎯 |
|
- Proximal Policy Optimization Fine Tuning 🤝 |
|
- Variations - Preference Model Pretraining 🤔 |
|
- Use Ranking Datasets Sentiment - Thumbs Up/Down, Distribution 📊 |
|
- Online Version Getting Feedback 💬 |
|
- OpenAI - InstructGPT - Humans generate LM Training Text 🔍 |
|
- DeepMind - Advantage Actor Critic Sparrow, GopherCite 🦜 |
|
- Reward Model Human Prefence Feedback 🏆 |
|
|
|
For more information on specific techniques and implementations, check out the following resources: |
|
- OpenAI's paper on [GPT-3](https://arxiv.org/abs/2005.14165) which details their Language Model Preparation approach |
|
- DeepMind's paper on [SAC](https://arxiv.org/abs/1801.01290) which describes the Advantage Actor Critic algorithm |
|
- OpenAI's paper on [Reward Learning](https://arxiv.org/abs/1810.06580) which explains their approach to training Reward Models |
|
- OpenAI's blog post on [GPT-3's fine-tuning process](https://openai.com/blog/fine-tuning-gpt-3/) |
|
|
|
""") |
|
|
|
demo.launch() |
