Llama guard data formatter example #337
Conversation
…ration cofigs explanation
| @@ -0,0 +1,98 @@ | |||
| # Finetuning Data Formatter | |||
|
|
|||
| The finetuning_data_formatter script provides classes and methods for formatting training data for finetuning a language model on a specific task. The main classes are: | |||
There was a problem hiding this comment.
"a language model" -> Llama Guard
There was a problem hiding this comment.
This isn't for all LMs we have
There was a problem hiding this comment.
Quick question, if someone wants to train llama guard from scratch on top of llama 2 7B, can they use this formatter as well?
| The finetuning_data_formatter script provides classes and methods for formatting training data for finetuning a language model on a specific task. The main classes are: | ||
| * `TrainingExample`: Represents a single example in the training data, consisting of a prompt, response, label (safe or unsafe), violated category codes, and an explanation. | ||
| * `Guidelines`: Defines the categories and their descriptions that will be used to evaluate the safety of the responses. | ||
| * `LlamaGuardPromptConfigs`: Configures how the prompt that will be given to the language model during finetuning should be formatted. |
| * `TrainingExample`: Represents a single example in the training data, consisting of a prompt, response, label (safe or unsafe), violated category codes, and an explanation. | ||
| * `Guidelines`: Defines the categories and their descriptions that will be used to evaluate the safety of the responses. | ||
| * `LlamaGuardPromptConfigs`: Configures how the prompt that will be given to the language model during finetuning should be formatted. | ||
| * `LlamaGuardGenerationConfigs`: Configures how the language model's response should be formatted. |
There was a problem hiding this comment.
I will omit this comment after this point, but in general recommend to replace all relevant instances of LM with LG in this doc :)
| - First line must read 'safe' or 'unsafe'. | ||
| - If unsafe, a second line must include a comma-separated list of violated categories. """, | ||
| should_include_category_descriptions=True, | ||
| should_shuffle_category_codes=False |
There was a problem hiding this comment.
Might recommend to set to True in the example here. May make it more robust. Hakan to comment
There was a problem hiding this comment.
@inanhkn changed the example to True already
| ) | ||
| ``` | ||
|
|
||
| Then, you need to configure the prompt that will be given to the language model during finetuning. You do this by creating an instance of the LlamaGuardPromptConfigs class and specifying the format string and other options. For example: |
There was a problem hiding this comment.
Might be worth pointing out that the finetuning and inference-time prompts should be the structurally the same for best performance
There was a problem hiding this comment.
In line with this, should the new prompts be structurally similar to the ones used by Llama Guard?
| explanation="The response contains personal information." | ||
| ), | ||
| TrainingExample( | ||
| prompt="Where do you live?", | ||
| response="I live in New York City.", | ||
| violated_category_codes=["O2"], | ||
| label="unsafe", | ||
| explanation="The response reveals the user's location." |
There was a problem hiding this comment.
Probably doesn't matter, but if the bot responds with "my name is john" or "I live in NYC", that doesn't appear to leak a person's info or location, but the bot's. Maybe "what is the name of the McDonald's manager at 123 Main street" and "where does Voltaire Strongfeld live" are questions that would elicit this info.
Not a big deal, because this is just an example, but there's a tiny chance it may confuse someone
There was a problem hiding this comment.
Overall super nit on this file: would add a line of space between each variable declaration for easier readingl/skimming
There was a problem hiding this comment.
Added space and changed the order of the code for better readibily
There was a problem hiding this comment.
just a couple of final comments that Beto can decide if changes are needed. @albertodepaola
What does this PR do?
Adds a simple example on how to use the data formatter script for fine tuning Llama Guard. Adds a readme explaining the steps in the script as well.
Testing
Running the script and checking that it generated valid prompts:
python src/llama_recipes/data/llama_guard/finetuning_data_formatter_example.pyOutput is as show in the file:
sample_formatted_data.json
Before submitting
Pull Request section?
to it if that's the case.
Thanks for contributing 🎉!