188 lines
10 KiB
Markdown
188 lines
10 KiB
Markdown
---
|
|
license: mit
|
|
language:
|
|
- en
|
|
library_name: transformers
|
|
inference: false
|
|
datasets:
|
|
- databricks/databricks-dolly-15k
|
|
---
|
|
# dolly-v2-12b Model Card
|
|
## Summary
|
|
|
|
Databricks' `dolly-v2-12b`, an instruction-following large language model trained on the Databricks machine learning platform
|
|
that is licensed for commercial use. Based on `pythia-12b`, Dolly is trained on ~15k instruction/response fine tuning records
|
|
[`databricks-dolly-15k`](https://github.com/databrickslabs/dolly/tree/master/data) generated
|
|
by Databricks employees in capability domains from the InstructGPT paper, including brainstorming, classification, closed QA, generation,
|
|
information extraction, open QA and summarization. `dolly-v2-12b` is not a state-of-the-art model, but does exhibit surprisingly
|
|
high quality instruction following behavior not characteristic of the foundation model on which it is based.
|
|
|
|
Dolly v2 is also available in these smaller models sizes:
|
|
|
|
* [dolly-v2-7b](https://huggingface.co/databricks/dolly-v2-7b), a 6.9 billion parameter based on `pythia-6.9b`
|
|
* [dolly-v2-3b](https://huggingface.co/databricks/dolly-v2-3b), a 2.8 billion parameter based on `pythia-2.8b`
|
|
|
|
Please refer to the [dolly GitHub repo](https://github.com/databrickslabs/dolly#getting-started-with-response-generation) for tips on
|
|
running inference for various GPU configurations.
|
|
|
|
**Owner**: Databricks, Inc.
|
|
|
|
## Model Overview
|
|
`dolly-v2-12b` is a 12 billion parameter causal language model created by [Databricks](https://databricks.com/) that is derived from
|
|
[EleutherAI's](https://www.eleuther.ai/) [Pythia-12b](https://huggingface.co/EleutherAI/pythia-12b) and fine-tuned
|
|
on a [~15K record instruction corpus](https://github.com/databrickslabs/dolly/tree/master/data) generated by Databricks employees and released under a permissive license (CC-BY-SA)
|
|
|
|
## Usage
|
|
|
|
To use the model with the `transformers` library on a machine with GPUs, first make sure you have the `transformers` and `accelerate` libraries installed.
|
|
In a Databricks notebook you could run:
|
|
|
|
```python
|
|
%pip install "accelerate>=0.16.0,<1" "transformers[torch]>=4.28.1,<5" "torch>=1.13.1,<2"
|
|
```
|
|
|
|
The instruction following pipeline can be loaded using the `pipeline` function as shown below. This loads a custom `InstructionTextGenerationPipeline`
|
|
found in the model repo [here](https://huggingface.co/databricks/dolly-v2-3b/blob/main/instruct_pipeline.py), which is why `trust_remote_code=True` is required.
|
|
Including `torch_dtype=torch.bfloat16` is generally recommended if this type is supported in order to reduce memory usage. It does not appear to impact output quality.
|
|
It is also fine to remove it if there is sufficient memory.
|
|
|
|
```python
|
|
import torch
|
|
from transformers import pipeline
|
|
|
|
generate_text = pipeline(model="databricks/dolly-v2-12b", torch_dtype=torch.bfloat16, trust_remote_code=True, device_map="auto")
|
|
```
|
|
|
|
You can then use the pipeline to answer instructions:
|
|
|
|
```python
|
|
res = generate_text("Explain to me the difference between nuclear fission and fusion.")
|
|
print(res[0]["generated_text"])
|
|
```
|
|
|
|
Alternatively, if you prefer to not use `trust_remote_code=True` you can download [instruct_pipeline.py](https://huggingface.co/databricks/dolly-v2-3b/blob/main/instruct_pipeline.py),
|
|
store it alongside your notebook, and construct the pipeline yourself from the loaded model and tokenizer:
|
|
|
|
```python
|
|
import torch
|
|
from instruct_pipeline import InstructionTextGenerationPipeline
|
|
from transformers import AutoModelForCausalLM, AutoTokenizer
|
|
|
|
tokenizer = AutoTokenizer.from_pretrained("databricks/dolly-v2-12b", padding_side="left")
|
|
model = AutoModelForCausalLM.from_pretrained("databricks/dolly-v2-12b", device_map="auto", torch_dtype=torch.bfloat16)
|
|
|
|
generate_text = InstructionTextGenerationPipeline(model=model, tokenizer=tokenizer)
|
|
```
|
|
|
|
### LangChain Usage
|
|
|
|
To use the pipeline with LangChain, you must set `return_full_text=True`, as LangChain expects the full text to be returned
|
|
and the default for the pipeline is to only return the new text.
|
|
|
|
```python
|
|
import torch
|
|
from transformers import pipeline
|
|
|
|
generate_text = pipeline(model="databricks/dolly-v2-12b", torch_dtype=torch.bfloat16,
|
|
trust_remote_code=True, device_map="auto", return_full_text=True)
|
|
```
|
|
|
|
You can create a prompt that either has only an instruction or has an instruction with context:
|
|
|
|
```python
|
|
from langchain import PromptTemplate, LLMChain
|
|
from langchain.llms import HuggingFacePipeline
|
|
|
|
# template for an instrution with no input
|
|
prompt = PromptTemplate(
|
|
input_variables=["instruction"],
|
|
template="{instruction}")
|
|
|
|
# template for an instruction with input
|
|
prompt_with_context = PromptTemplate(
|
|
input_variables=["instruction", "context"],
|
|
template="{instruction}\n\nInput:\n{context}")
|
|
|
|
hf_pipeline = HuggingFacePipeline(pipeline=generate_text)
|
|
|
|
llm_chain = LLMChain(llm=hf_pipeline, prompt=prompt)
|
|
llm_context_chain = LLMChain(llm=hf_pipeline, prompt=prompt_with_context)
|
|
```
|
|
|
|
Example predicting using a simple instruction:
|
|
|
|
```python
|
|
print(llm_chain.predict(instruction="Explain to me the difference between nuclear fission and fusion.").lstrip())
|
|
```
|
|
|
|
Example predicting using an instruction with context:
|
|
|
|
```python
|
|
context = """George Washington (February 22, 1732[b] - December 14, 1799) was an American military officer, statesman,
|
|
and Founding Father who served as the first president of the United States from 1789 to 1797."""
|
|
|
|
print(llm_context_chain.predict(instruction="When was George Washington president?", context=context).lstrip())
|
|
```
|
|
|
|
|
|
## Known Limitations
|
|
|
|
### Performance Limitations
|
|
**`dolly-v2-12b` is not a state-of-the-art generative language model** and, though quantitative benchmarking is ongoing, is not designed to perform
|
|
competitively with more modern model architectures or models subject to larger pretraining corpuses.
|
|
|
|
The Dolly model family is under active development, and so any list of shortcomings is unlikely to be exhaustive, but we include known limitations and misfires here as a means to document and share our preliminary findings with the community.
|
|
In particular, `dolly-v2-12b` struggles with: syntactically complex prompts, programming problems, mathematical operations, factual errors,
|
|
dates and times, open-ended question answering, hallucination, enumerating lists of specific length, stylistic mimicry, having a sense of humor, etc.
|
|
Moreover, we find that `dolly-v2-12b` does not have some capabilities, such as well-formatted letter writing, present in the original model.
|
|
|
|
### Dataset Limitations
|
|
Like all language models, `dolly-v2-12b` reflects the content and limitations of its training corpuses.
|
|
|
|
- **The Pile**: GPT-J's pre-training corpus contains content mostly collected from the public internet, and like most web-scale datasets,
|
|
it contains content many users would find objectionable. As such, the model is likely to reflect these shortcomings, potentially overtly
|
|
in the case it is explicitly asked to produce objectionable content, and sometimes subtly, as in the case of biased or harmful implicit
|
|
associations.
|
|
|
|
- **`databricks-dolly-15k`**: The training data on which `dolly-v2-12b` is instruction tuned represents natural language instructions generated
|
|
by Databricks employees during a period spanning March and April 2023 and includes passages from Wikipedia as references passages
|
|
for instruction categories like closed QA and summarization. To our knowledge it does not contain obscenity, intellectual property or
|
|
personally identifying information about non-public figures, but it may contain typos and factual errors.
|
|
The dataset may also reflect biases found in Wikipedia. Finally, the dataset likely reflects
|
|
the interests and semantic choices of Databricks employees, a demographic which is not representative of the global population at large.
|
|
|
|
Databricks is committed to ongoing research and development efforts to develop helpful, honest and harmless AI technologies that
|
|
maximize the potential of all individuals and organizations.
|
|
|
|
### Benchmark Metrics
|
|
|
|
Below you'll find various models benchmark performance on the [EleutherAI LLM Evaluation Harness](https://github.com/EleutherAI/lm-evaluation-harness);
|
|
model results are sorted by geometric mean to produce an intelligible ordering. As outlined above, these results demonstrate that `dolly-v2-12b` is not state of the art,
|
|
and in fact underperforms `dolly-v1-6b` in some evaluation benchmarks. We believe this owes to the composition and size of the underlying fine tuning datasets,
|
|
but a robust statement as to the sources of these variations requires further study.
|
|
|
|
| model | openbookqa | arc_easy | winogrande | hellaswag | arc_challenge | piqa | boolq | gmean |
|
|
| --------------------------------- | ------------ | ---------- | ------------ | ----------- | --------------- | -------- | -------- | ---------|
|
|
| EleutherAI/pythia-2.8b | 0.348 | 0.585859 | 0.589582 | 0.591217 | 0.323379 | 0.73395 | 0.638226 | 0.523431 |
|
|
| EleutherAI/pythia-6.9b | 0.368 | 0.604798 | 0.608524 | 0.631548 | 0.343857 | 0.761153 | 0.6263 | 0.543567 |
|
|
| databricks/dolly-v2-3b | 0.384 | 0.611532 | 0.589582 | 0.650767 | 0.370307 | 0.742655 | 0.575535 | 0.544886 |
|
|
| EleutherAI/pythia-12b | 0.364 | 0.627104 | 0.636148 | 0.668094 | 0.346416 | 0.760065 | 0.673394 | 0.559676 |
|
|
| EleutherAI/gpt-j-6B | 0.382 | 0.621633 | 0.651144 | 0.662617 | 0.363481 | 0.761153 | 0.655963 | 0.565936 |
|
|
| databricks/dolly-v2-12b | 0.408 | 0.63931 | 0.616417 | 0.707927 | 0.388225 | 0.757889 | 0.568196 | 0.56781 |
|
|
| databricks/dolly-v2-7b | 0.392 | 0.633838 | 0.607735 | 0.686517 | 0.406997 | 0.750816 | 0.644037 | 0.573487 |
|
|
| databricks/dolly-v1-6b | 0.41 | 0.62963 | 0.643252 | 0.676758 | 0.384812 | 0.773667 | 0.687768 | 0.583431 |
|
|
| EleutherAI/gpt-neox-20b | 0.402 | 0.683923 | 0.656669 | 0.7142 | 0.408703 | 0.784004 | 0.695413 | 0.602236 |
|
|
|
|
# Citation
|
|
|
|
```
|
|
@online{DatabricksBlog2023DollyV2,
|
|
author = {Mike Conover and Matt Hayes and Ankit Mathur and Jianwei Xie and Jun Wan and Sam Shah and Ali Ghodsi and Patrick Wendell and Matei Zaharia and Reynold Xin},
|
|
title = {Free Dolly: Introducing the World's First Truly Open Instruction-Tuned LLM},
|
|
year = {2023},
|
|
url = {https://www.databricks.com/blog/2023/04/12/dolly-first-open-commercially-viable-instruction-tuned-llm},
|
|
urldate = {2023-06-30}
|
|
}
|
|
```
|
|
|
|
# Happy Hacking! |