pretty_name: "GT-HarmBench Deont Adapter Eval"
license: mit
viewer: true
tags:
- evaluation
- game-theory
- safety
- alignment
- text
- csv
- jsonl
configs: - config_name: summary_table
default: true
data_files: "**/*.csv" - config_name: per_example_records
data_files: "**/*.jsonl"
GT-HarmBench Deont Adapter Eval
This dataset repo contains outputs from a 50-scenario GT-HarmBench evaluation run for the adapter:agentic-moral-alignment/qwen35-9b__ipd_str_tft__deont__native_tool__r1
Run settings:
- Base model:
unsloth/Qwen3.5-9B - Quantization:
no_load_in_4bit: true - Prompt mode:
bare - Reasoning mode:
native - Tool use:
true - Reasoning length hint:
100 - Max new tokens:
512 - Max scenarios:
50 - Seed:
100
Files at repo root:
The README YAML frontmatter defines two dataset viewer configurations:
summary_table: loads the root CSV.per_example_records: loads the root JSONL.
The summary JSON is included as an artifact, but it is not added as a viewer config because it is a single aggregate object rather than a row-oriented data file.