Updates

README.md

@@ -87,14 +87,12 @@ widget:
 
 #  Table of Contents
 
-1. [Model Summary](#model=summary)
+1. [Model Summary](#model-summary)
 2. [Use](#use)
-3. [Bias, Risks, and Limitations](#bias-risks-and-limitations)
-4. [Training Details](#training-details)
+3. [Limitations](#limitations)
+4. [Training](#training)
 5. [Evaluation](#evaluation)
-6. [Environmental Impact](#environmental-impact)
 7. [Citation](#citation)
-9. [How To Get Started With the Model](#how-to-get-started-with-the-model)
 
 # Model Summary
 
@@ -103,6 +101,7 @@ widget:
 - **Repository:** [bigscience-workshop/xmtf](https://github.com/bigscience-workshop/xmtf)
 - **Paper:** [TODO]
 - **Point of Contact:** [Niklas Muennighoff](mailto:niklas@hf.co)
+- **Languages:** Refer to [BLOOM](https://huggingface.co/bigscience/bloom) for pretraining & [xP3](https://huggingface.co/bigscience/xP3) for finetuning language proportions. It understands both pretraining & finetuning languages.
 - **BLOOMZ & mT0 Model Family:**
 |Name|Explanation|
 |----|-----------|
@@ -129,40 +128,117 @@ widget:
 |[mt0-xxl-mt](https://huggingface.co/bigscience/mt0-xxl-mt)|13B parameter multitask finetuned version of [mt5-xxl](https://huggingface.co/google/mt5-xxl) on [xP3](https://huggingface.co/bigscience/xP3) & [xP3mt](https://huggingface.co/bigscience/xP3mt). **Better than [mt0-xxl](https://huggingface.co/bigscience/mt0-xxl) when prompting in non-English**|
 |||
 |[mt0-xxl-p3](https://huggingface.co/bigscience/mt0-xxl-p3)| 13B parameter multitask finetuned version of [mt5-xxl](https://huggingface.co/google/mt5-xxl) on [P3](https://huggingface.co/bigscience/P3). **Released for research purposes, performance is inferior to [mt0-xxl](https://huggingface.co/bigscience/mt0-xxl)**|
-|----|-----------|
 
+# Use
+
+## Intended use
+
+We recommend using the model to perform tasks expressed in natural language. For example, given the prompt "*Translate to English: Je t’aime.*", the model will most likely answer "*I love you.*" Some other prompt ideas from our paper: 
+- 一个传奇的开端，一个不灭的神话，这不仅仅是一部电影，而是作为一个走进新时代的标签，永远彪炳史册。你认为这句话的立场是赞扬、中立还是批评?
+- Suggest at least five related search terms to "Mạng neural nhân tạo".
+- Write a fairy tale about a troll saving a princess from a dangerous dragon. The fairy tale is a masterpiece that has achieved praise worldwide and its moral is "Heroes Come in All Shapes and Sizes". Story (in Spanish):
+- Explain in a sentence in Telugu what is backpropagation in neural networks.
 
-# Intended uses
+**Feel free to share your generations in the Community tab!**
 
-You can use the models to perform inference on tasks by specifying your query in natural language, and the models will generate a prediction. For instance, you can ask *"Translate this to Chinese: Je t'aime."*, and the model will hopefully generate *"我爱你"*.
+## How to use
 
-# How to use
+### CPU
 
-Here is how to use the model in PyTorch:
+<details>
+<summary> Click to expand </summary>
 
-TODO: Better code with auto-precision?
 ```python
-from transformers import AutoTokenizer, AutoModelForCausalLM
+# pip install -q transformers
+from transformers import AutoModelForCausalLM, AutoTokenizer
 
-tokenizer = AutoTokenizer.from_pretrained("bigscience/bloomz-560m")
-model = AutoModelForCausalLM.from_pretrained("bigscience/bloomz-560m")
+checkpoint = "bigscience/bloomz"
 
-inputs = tokenizer.encode("Is this review positive or negative? Review: this is the best cast iron skillet you will ever buy", return_tensors="pt")
+tokenizer = AutoTokenizer.from_pretrained(checkpoint)
+model = AutoModelForCausalLM.from_pretrained(checkpoint)
+
+inputs = tokenizer.encode("Translate to English: Je t’aime.", return_tensors="pt")
 outputs = model.generate(inputs)
 print(tokenizer.decode(outputs[0]))
 ```
 
-To use another checkpoint, replace the path in `AutoTokenizer` and `AutoModelForCausalLM`.
+</details>
+
+### GPU
+
+<details>
+<summary> Click to expand </summary>
+
+```python
+# pip install -q transformers accelerate
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+checkpoint = "bigscience/bloomz"
 
-**Note: 176B models are trained with bfloat16, while smaller models are trained with fp16. We recommend using the same precision type or fp32 at inference**
+tokenizer = AutoTokenizer.from_pretrained(checkpoint)
+model = AutoModelForCausalLM.from_pretrained(checkpoint, torch_dtype="auto", device_map="auto")
+
+inputs = tokenizer.encode("Translate to English: Je t’aime.", return_tensors="pt").to("cuda")
+outputs = model.generate(inputs)
+print(tokenizer.decode(outputs[0]))
+```
+
+</details>
+
+### GPU in 8bit
+
+<details>
+<summary> Click to expand </summary>
+
+```python
+# pip install -q transformers accelerate bitsandbytes
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+checkpoint = "bigscience/bloomz"
+
+tokenizer = AutoTokenizer.from_pretrained(checkpoint)
+model = AutoModelForCausalLM.from_pretrained(checkpoint, device_map="auto", load_in_8bit=True)
+
+inputs = tokenizer.encode("Translate to English: Je t’aime.", return_tensors="pt").to("cuda")
+outputs = model.generate(inputs)
+print(tokenizer.decode(outputs[0]))
+```
+
+</details>
 
 # Limitations
 
-- Large model size may require large computational resources
-- High performance variance depending on the prompt
+The performance may vary depending on the prompt. For BLOOMZ models, we recommend making it very clear when the input stops to avoid the model trying to continue it. For example, the prompt "*Translate to English: Je t'aime*" without the full stop (.) at the end, may result in the model trying to continue the French sentence. Better prompts are e.g. "*Translate to English: Je t'aime.*", "*Translate to English: Je t'aime. Translation:*" "*What is "Je t'aime." in English?*", where it is clear for the model when it should answer. Further, we recommend providing the model as much context as possible. For example, if you want it to answer in Telugu, then tell the model, e.g. "*Explain in a sentence in Telugu what is backpropagation in neural networks.*".
+
+# Training
+
+## Model
+
+- Architecture: Same as [bloom](https://huggingface.co/bigscience/bloom), also refer to the `config.json` file
+- Finetuning steps: 498
+- Finetuning tokens: 2.09 billion
+- Finetuning layout: 72x pipeline parallel, 1x tensor parallel, 4x data parallel
+- Precision: bfloat16
+
+## Hardware
+
+- 288 A100 80GB GPUs (36 nodes)
+- 8 GPUs per node using NVLink 4 inter-gpu connects, 4 OmniPath links
+- NCCL-communications network: a fully dedicated subnet
+- AMD CPUs with 512GB memory per node
+
+## Software
+
+- [Megatron-DeepSpeed](https://github.com/bigscience-workshop/Megatron-DeepSpeed)
+- [DeepSpeed](https://github.com/microsoft/DeepSpeed))
+- [PyTorch](https://github.com/pytorch/pytorch) (pytorch-1.11 w/ CUDA-11.5)
+- [apex](https://github.com/NVIDIA/apex)
+
+# Evaluation
 
-# BibTeX entry and citation info
+We refer to Table 7 from our paper [TODO LINK].
 
+# Citation
 ```bibtex
 TODO
 ```