File size: 3,798 Bytes
006313a
 
 
 
 
 
 
 
 
 
42ec4d1
 
 
 
4cbd52c
08cf7ac
 
375f128
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
08cf7ac
 
 
 
 
 
 
 
b25009a
 
 
28a2066
b25009a
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
---
license: bigscience-openrail-m
datasets:
- lvwerra/stack-exchange-paired
language:
- en
tags:
- trl
- transformers
- rlhf
---

# Stack-Llama-2

[DPO](https://github.com/eric-mitchell/direct-preference-optimization) fine-tuned [Llama-2 7B model](https://huggingface.co/meta-llama/Llama-2-7b). The model is designed to generate human-like responses to questions in Stack Exchange domains of programming, mathematics, physics, and more. For more info check out the [blog post](https://huggingface.co/blog/dpo-trl) and github [example](https://github.com/lvwerra/trl/tree/main/examples/research_projects/stack_llama_2/scripts).


## Uses

### Direct Use
- Long-form question-answering on topics of programming, mathematics, and physics
- Demonstrating a Large Language Model's ability to follow target behavior of generating answers to a question that would be highly rated on [Stack Exchange](https://stackexchange.com).

### Out of Scope Use
- Replacing human expertise


## Bias, Risks, and Limitations
- Inherits bias, risks, and limitations from the LLaMA model, as described in the [LLaMA Model Card Bias Evaluation](https://github.com/facebookresearch/llama/blob/main/MODEL_CARD.md#quantitative-analysis) and [Ethical Considerations](https://github.com/facebookresearch/llama/blob/main/MODEL_CARD.md#ethical-considerations). 
- Retains biases present in the Stack Exchange dataset. Per the [latest developer survey for Stack Overflow](https://survey.stackoverflow.co/2022/),
which constitutes a significant part of the StackExchange data,
most users who answered the survey identified themselves as [White or European, men, between 25 and 34 years old, and based in the US (with a significant part of responders from India).](https://survey.stackoverflow.co/2022/#developer-profile-demographics)
- May generate answers that are incorrect or misleading.
- May copy answers from the training data verbatim.
- May generate language that is hateful or promotes discrimination ([example](https://huggingface.co/trl-lib/llama-7b-se-rl-peft/discussions/7#64376083369f6f907f5bfe4c)).
- May generate language that is offensive to direct or indirect users or to people or groups mentioned.


### Recommendations
- Answers should be validated through the use of external sources.
- Disparities between the data contributors and the direct and indirect users of the technology should inform developers in assessing what constitutes an appropriate use case.
- Further research is needed to attribute model generations to sources in the training data, especially in cases where the model copies answers from the training data.  


## Training Details

### Training Data

Original datasets are described in [the LLaMA Model Card](https://github.com/facebookresearch/llama/blob/main/MODEL_CARD.md#training-dataset).
Fine-tuning datasets for this model are based on [Stack Exchange Paired](https://huggingface.co/datasets/lvwerra/stack-exchange-paired), which consists of questions and answers from various domains in Stack Exchange, such as programming, mathematics, physics, and more. Specifically:

**Traditional Fine-tuning:** [https://huggingface.co/datasets/lvwerra/stack-exchange-paired/tree/main/data/finetune](https://huggingface.co/datasets/lvwerra/stack-exchange-paired/tree/main/data/finetune)

**DPO Training:** [https://huggingface.co/datasets/lvwerra/stack-exchange-paired/tree/main/data/rl](https://huggingface.co/datasets/lvwerra/stack-exchange-paired/tree/main/data/rl)

### Training Procedure
The model was first fine-tuned on the Stack Exchange question and answer pairs and then fine-tuned via the DPO training procedure using the SFT model as the reference model.   It is trained to respond to prompts with the following prompt template:

```
Question: <Query> 

Answer: <Response>
```