Merge branch 'main' of https://github.com/nomic-ai/gpt4all into main

README.md

@@ -1,13 +1,50 @@
-# gpt4all
+<h1 align="center">GPT4All</h1>
+<p align="center">Demo, data and code to train an assistant-style large language model</p>
+
+<p align="center">
+<a href="https://s3.amazonaws.com/static.nomic.ai/gpt4all/2023_GPT4All_Technical_Report.pdf">:green_book: Technical Report</a>
+</p>
 
 
 
-# Setup
+![gpt4all-lora-demo](https://user-images.githubusercontent.com/13879686/228352356-de66ca7a-df70-474e-b929-2e3656165051.gif)
+
+
+# Try it yourself
+You can download pre-compiled LLaMa C++ Interactive Chat binaries here:
+- [OSX](https://s3.amazonaws.com/static.nomic.ai/gpt4all/models/gpt4all-lora-quantized-OSX-m1)
+- [Intel/Windows]()
+
+and the model
+- [gpt4all-quantized](https://s3.amazonaws.com/static.nomic.ai/gpt4all/models/gpt4all-lora-quantized.bin)
+
+Place the binary and quantized model in the same directory and start chatting!
+
+To compile for custom hardware, see our fork of the [Alpaca C++](https://github.com/zanussbaum/alpaca.cpp) repo.
+
+
+# Reproducibility
+
+Trained LoRa Weights:
+- gpt4all-lora:  https://huggingface.co/nomic-ai/gpt4all-lora
+- gpt4all-lora-epoch-2 https://huggingface.co/nomic-ai/gpt4all-lora-epoch-2
+
+Raw Data:
+- [Training Data Without P3](https://s3.amazonaws.com/static.nomic.ai/gpt4all/2022_03_27/gpt4all_curated_data_without_p3_2022_03_27.tar.gz)
+- [Full Dataset with P3](https://s3.amazonaws.com/static.nomic.ai/gpt4all/2022_03_27/gpt4all_curated_data_full_2022_03_27.tar.gz)
+
+We are not distributing a LLaMa 7B checkpoint.
+
+You can reproduce our trained model by doing the following:
+
+## Setup
 
 Clone the repo
 
 `git clone --recurse-submodules git@github.com:nomic-ai/gpt4all.git`
 
+`git submodule configure && git submodule update`
+
 Setup the environment
 
 ```
@@ -20,12 +57,29 @@ cd ../peft
 pip install -e .
 ```
 
+## Training
+
+```bash
+accelerate launch --dynamo_backend=inductor --num_processes=8 --num_machines=1 --machine_rank=0 --deepspeed_multinode_launcher standard --mixed_precision=bf16  --use_deepspeed --deepspeed_config_file=configs/deepspeed/ds_config.json train.py --config configs/train/finetune-7b.yaml
+```
 
 ## Generate
 
-`python generate.py --config configs/generate/generate.yaml --prompt "Write a script to reverse a string in Python`
+```bash
+python generate.py --config configs/generate/generate.yaml --prompt "Write a script to reverse a string in Python
+```
 
 
-## Train
 
-`accelerate launch --dynamo_backend=inductor --num_processes=8 --num_machines=1 --machine_rank=0 --deepspeed_multinode_launcher standard --mixed_precision=bf16  --use_deepspeed --deepspeed_config_file=configs/deepspeed/ds_config.json train.py --config configs/train/finetune-7b.yaml`
+If you utilize this reposistory, models or data in a downstream project, please consider citing it with:
+```
+@misc{gpt4all,
+  author = {Yuvanesh Anand and Zach Nussbaum and Brandon Duderstadt and Benjamin Schmidt and Andriy Mulyar},
+  title = {GPT4All: Training an Assistant-style Chatbot with Large Scale Data Distillation from GPT-3.5-Turbo},
+  year = {2023},
+  publisher = {GitHub},
+  journal = {GitHub repository},
+  howpublished = {\url{https://github.com/nomic-ai/gpt4all}},
+}
+```
+

TRAINING_LOG.md

@@ -0,0 +1,233 @@
+# Intro
+
+This is a training log for both the LoRa and full model training we underwent
+
+## Inspiration
+
+
+
+## Initial Experiment
+
+We train an initial LoRa model on ~700k examples including data from P3/BLOOM, StackOverflow, and unified_chip2.
+We trained using 8 x A100 80GB GPUs.
+
+We used the initial parameters:
+
+| Hyperparameter | Value |
+|----------------|-------|
+| Per Device BS  | 4     |
+| Global BS      | 32    |
+| Learning rate  | 5e-5  |
+| Epochs         | 1     |
+| Max length     | 1024  |
+| Weight decay   | 0     |
+| Warmup Steps   | 100   |
+
+We randomly shuffle and set aside %5 of the data for validation.
+
+We had an initial bug in logging the training loss but we noticed a decrease in validation loss.
+
+![](figs/first_lora.png)
+
+
+During generation, we noticed that our model seemed to never generate an `eos` token. We realized that this was due to the fact that we set
+the tokenizer `pad` token equal to the `eos` token and most likely never learned to "stop". However, generations seemed to be working
+well, even with greedy generation. 
+
+For example, when prompted with `Generate a python script to reverse a string.`, the model outputs valid markdown and a working script
+
+    Here's a Python script to reverse a string:
+
+    ```python
+    def reverse_string(string):
+        return ''.join(reversed(string))
+    ```
+
+    This function takes a string as an argument and returns the reversed string. The `reversed()` function reverses the string and returns a list of characters. The `join()` function joins the list of characters into a string.
+
+    You can use this function like this:
+
+    ```python
+    string = 'hello world'
+    print(reverse_string(string))
+    ```
+
+    This will output `'olleh dna hwl'`.
+
+    Note that this function does not handle strings with spaces or other special characters. If you need to handle those cases, you can use a regular expression to split the string into words and then reverse each word.
+
+    ```python
+    import re
+
+    def reverse_string(string):
+        words = re.split(' ', string)
+        return ''.join(reversed(word) for word in words)
+    ```
+
+    This function takes a string as an argument and returns the reversed string. The `re.split()` function splits the string into words using a space as the delimiter. The `reversed()` function reverses each word and returns a list of characters. The `join()` function joins the list of characters into a string.
+
+    You can use this function like this:
+
+    ```python
+    string = 'hello world'
+    print(reverse_string(string))
+    ```
+
+    This will output `'olleh dna hwl'`.
+
+    Note that this function also handles strings with spaces or other special characters.
+
+
+    Answer: Here's a Python script to reverse a string:
+
+    ```python
+    def reverse_string(string):
+        return ''.join(reversed(string))
+    ```
+
+    This function takes a string as an argument and returns the reversed string. The `reversed()` function reverses the string and returns a list of characters. The `join()` function joins the list of characters into a string.
+
+    You can use this function like this:
+
+    ```python
+    string = 'hello world'
+    print(reverse_string(string))
+    ```
+
+    This will output `'olleh dna hwl'`.
+
+    Note that this function does not handle strings with
+
+although it duplicates and continues generating past answering the full question.
+
+The base model doesn't answer the question and hallucinates
+```
+Write a python script to reverse a string. The string should be reversed in place.
+The string should be reversed in place.
+The string should be reversed in place. The string should be reversed in place.
+The string should be reversed in place. The string should be reversed in place. The string should be reversed in place.
+The string should be reversed in place. The string should be reversed in place. The string should be reversed in place. The string should be reversed in place.
+The string should be reversed in place. The string should be reversed in place. The string should be reversed in place. The string should be reversed in place. The string should be reversed in place.
+The string should be reversed in place. The string should be reversed in place. The string should be reversed in place. The string should be reversed in place. The string should be reversed in place. The string should be reversed in place.
+The string should be reversed in place. The string should be reversed in place. The string should be reversed in place. The string should be reversed in place. The string should be reversed in place. The string should be reversed in place. The string should be reversed in place.
+The string should be reversed in place. The string should be reversed in place. The string should be reversed in place. The string should be reversed in place. The string should be reversed in place. The string should be reversed in place. The string should be reversed in place.
+The string should be reversed in place. The string should be reversed in place. The string should be reversed in place. The string should be reversed in place. The string should be reversed in place. The string should be reversed in place. The string should be reversed in place.
+The string should be reversed in place. The string should be reversed in place. The string should be reversed in place. The string should be reversed in place. The string should be reversed in place. The string should be reversed in place. The string should be reversed in place.
+The string should be reversed in place. The string should be reversed in place. The string should be reversed in place. The string should be reversed in place. The string should be
+```
+
+
+## EOS and Accidental Duplication
+
+Seeing as the model never stopped, we experimented with training a model with a separate token for `eos` and `pad`. Since we couldn't find a `pad` token present in the vocabulary, we added one to the tokenizer and expanded the embedding size of the model (from 32,000). In theory, we could have expanded the embedding size by a factor of 64 to improve throughput and performance, as [noted here](https://twitter.com/ctnzr/status/1623758178587648000?s=20).
+For every sequence, we appended an `eos` token in hopes the model would learn to exit. 
+
+We successfully trained a model using the same parameters as before ![](figs/duplicate_loss.png)
+
+During generation, our model exited early even with greedy generations 
+
+    You can use the `reversed()` function to reverse a string in Python. Here's an example:
+
+    ```python
+    string = 'hello world'
+    reversed_string = reversed(string)
+    print(reversed_string)
+    ```
+
+    Output:
+
+    ```
+    world hello
+    ```
+
+For a harder prompt where we try to trick the model with comments
+
+```python
+#this code prints a string reversed
+my_string = "hello how are you"
+print(len(my_string))
+
+
+My code above does not work. Can you help me?
+```
+
+The model correctly answers
+
+    The code you provided does not print a string reversed. It prints the length of the string "hello how are you".
+
+
+We realized that we had two bugs however:
+- We accidentally duplicated data and effectively trained for 2 epochs instead of 1
+- We added an eos token to every sequence, even those that we truncated (e.g. long code that exceeds the 1024).
+
+## Conditonal EOS and 1 Epoch
+
+Using the same parameters, we then trained a model using a "conditional" eos token where we only add an `eos` when the inputs are less than the maximum sequence length for one epoch.
+
+Our training and validation loss looks fairly normal
+
+![](figs/single_epoch.png)
+
+However, we found generations to be slightly poorer. For the same prompt of `Generate a python script to reverse a string.`, the model generates
+
+    Write a python script to reverse a string.
+    Here is a python script to reverse a string:
+    import string
+
+    def reverse_string(string):
+        return string[::-1]
+
+    print(reverse_string('hello world'))
+
+    Output:
+    world hello
+
+The prompt 
+```python
+#this code prints a string reversed
+my_string = "hello how are you"
+print(len(my_string))
+
+
+My code above does not work. Can you help me?
+```
+
+does not generate any text.
+
+
+And the prompt `"Generate a python script to make a get request to an api endpoint."` generates
+
+    I'm sorry, I cannot provide a specific answer to this question as it requires more context and details about the API endpoint and the specific task you are trying to accomplish. Can you please provide more information?
+
+
+## Multi Epoch and Full Model Training
+
+We decided to remove the entire Bigscience/P3 subset from the final training dataset due to data diversity considerations. 
+P3 contains many homogeneous prompts which produce short and homogeneous responses from GPT-3.5-Turbo. 
+The final dataset is ~400k examples.
+
+We train a LoRa model using the parameters 
+
+| Hyperparameter | Value |
+|----------------|-------|
+| Per Device BS  | 4     |
+| Global BS      | 32    |
+| Learning rate  | 5e-5  |
+| Epochs         | 4     |
+| Max length     | 1024  |
+| Weight decay   | 0     |
+| Warmup Steps   | 100   |
+
+
+We additionally train a full model 
+| Hyperparameter | Value |
+|----------------|-------|
+| Per Device BS  | 32    |
+| Global BS      | 256   |
+| Learning rate  | 5e-5  |
+| Epochs         | 2     |
+| Max length     | 1024  |
+| Weight decay   | 0     |
+| Warmup Steps   | 100   |
+
+Taking inspiration from [the Alpaca Repo](https://github.com/tatsu-lab/stanford_alpaca), we roughly scale the learning rate by `sqrt(k)`, where `k` is the increase in batch size, where Alpaca used a batch size of 128 and learning rate of 2e-5.

configs/generate/generate.yaml

@@ -1,8 +1,8 @@
 # model/tokenizer
-model_name: "zpn/llama-7b"
+model_name: # REPLACE HERE with the base llama model 
-tokenizer_name: "zpn/llama-7b"
+tokenizer_name: # REPLACE HERE with the llama tokenizer
 lora: true
-lora_path: "nomic-ai/vicuna-lora-1024"
+lora_path: "nomic-ai/gpt4all-lora"
 
 max_new_tokens: 512
 temperature: 0

configs/generate/generate_llama.yaml

@@ -1,6 +1,6 @@
 # model/tokenizer
-model_name: "zpn/llama-7b"
+model_name: # REPLACE WITH LLAMA MODEL NAME
-tokenizer_name: "zpn/llama-7b"
+tokenizer_name: # REPLACE WITH LLAMA MODEL NAME
 
 
 max_new_tokens: 512