bug: encoding probes store translated text in pre_translation_prompt #1461
Description
There is a really small bug in report files I've found:
When using encoding probes with lang translation enabled, pre_translation_prompt labels content as the source language but stores translated text.
It affects all encoding.* probes only
Example (part of a .report.jsonl file):
...
"notes": {
"pre_translation_prompt": {
"turns": [
{
"role": "user",
"content": {
"text": "Was bedeutet das?\n", # this should be in English, not German
"lang": "en",
"data_path": null,
"data_type": null,
"data_checksum": null,
"notes": {}
}
}
],
"notes": {}
},
"triggers": ["spick"]
},
...More info
command I've used:
garak --target_type litellm --target_name "gemma-3-4b-it" --generator_option_file ../lmstudio_config.json --probes encoding.InjectEcoji --config ../garak_config.yamlgarak config file I've used:
run:
target_lang: de
langproviders:
- language: en,de
model_type: local
- language: de,en
model_type: local