Add audio NIM model and audio probes #1163
Conversation
erickgalinkin
force-pushed
the
audio-probes
branch
from
65c2dae to
633969e
Compare
erickgalinkin
force-pushed
the
audio-probes
branch
from
633969e to
3a5282a
Compare
| audio_achilles_data_dir = ( | ||
| _config.transient.cache_dir / "data" / "audio_achilles" | ||
| ) |
There was a problem hiding this comment.
prefer load using the data pattern:
| audio_achilles_data_dir = ( | |
| _config.transient.cache_dir / "data" / "audio_achilles" | |
| ) | |
| from garak.data import path as data_path | |
| audio_achilles_data_dir = ( | |
| data_path / "audio_achilles" | |
| ) |
This will move the location expected to $XDG_DATA_HOME/garak/data/audio_achilles.
This can serve as a good example of how a user can bring their own data files. The data_path should be treated as read only by garak and can be provided by the project installation or overridden/extended by the user. The _config.transient.cache_dir paths should not be user provided files as garak expected to manage files in that path.
| def write_audio_to_file(audio_data, file_path, sampling_rate): | ||
| """Writes audio data to a file. | ||
|
|
||
| Args: | ||
| audio_data: A 1D numpy array containing the audio data. | ||
| file_path: The path to the output audio file. | ||
| sampling_rate: The sampling rate of the audio data. | ||
| """ | ||
| sf.write(file_path, audio_data, sampling_rate) | ||
|
|
||
| import soundfile as sf | ||
| from datasets import load_dataset | ||
|
|
||
| os.makedirs(audio_achilles_data_dir) | ||
| dataset = load_dataset("garak-llm/audio_achilles_heel") | ||
| for item in dataset["train"]: | ||
| audio_data = item["audio"]["array"] | ||
| sampling_rate = item["audio"]["sampling_rate"] | ||
| file_path = str(audio_achilles_data_dir / item["audio"]["path"]) | ||
| write_audio_to_file(audio_data, file_path, sampling_rate) |
There was a problem hiding this comment.
This explains the reasoning for searching the cache_dir path, I assume this is meant to mirror the visual_jailbreak probe. In the visual_jailbreak probe however the cache_dir usage is due to the direct download of data into a cached location, is this really analogous to that?
I theory the the structure of the tests sent by this probe could allow user provided audio files in data however connecting the content of that file to an expected mitigation seems to be missing a link in the chain to allow/enable inspection of the data the detector result is based on.
Adds support for audio probes using [the datasets I have somewhere] and multimodal NIM.