Fairnsss Scoring Keywords & Max Tokens

[danielz02]

Describe the bug (Initially discovered and reported by UT Austin's VITA group. ) For models with verbose outputs, the max_tokens=20 setting in the fairness perspective is too small. This leads to truncated prediction and incorrect result parsing. In addition, fairness scoring metrics are lacking certain keywords. This might also be a problem when the model is not instruction-following.

To Reproduce

 dt-run +model_config=hf ++model_config.model=hf/meta-llama/Llama-2-13b-chat-hf +fairness=few_shot_tr_br_1.0
 cat results/fairness/results/hf/meta-llama/Llama-2-13b-chat-hf/few_shot_tr_br_1.0.json | jq

Traceback (most recent call last):
  File "/home/danielz/PycharmProjects/DecodingTrust/src/dt/perspectives/fairness/score_calculation_script.py", line 193, in <module>
    aggregate_current_scores(fairness_args)
  File "/home/danielz/PycharmProjects/DecodingTrust/src/dt/perspectives/fairness/score_calculation_script.py", line 142, in aggregate_current_scores
    m_dpd_ = demographic_parity_difference(label, prediction, sensitive_features=sensitive_attr)
  File "/home/danielz/anaconda3/lib/python3.9/site-packages/fairlearn/metrics/_fairness_metrics.py", line 45, in demographic_parity_difference
    sel_rate = MetricFrame(
  File "/home/danielz/anaconda3/lib/python3.9/site-packages/fairlearn/metrics/_metric_frame.py", line 103, in compatible_metric_frame_init
    new_metric_frame_init(self, **metric_arg_dict, **positional_dict, **kwargs)
  File "/home/danielz/anaconda3/lib/python3.9/site-packages/fairlearn/metrics/_metric_frame.py", line 309, in __init__
    sf_list = self._process_features("sensitive_feature_", sensitive_features, y_t)
  File "/home/danielz/anaconda3/lib/python3.9/site-packages/fairlearn/metrics/_metric_frame.py", line 856, in _process_features
    raise ValueError(_TOO_MANY_FEATURE_DIMS)
ValueError: Feature array has too many dimensions

Example Outputs

• Based on the attributes you provided, here is a prediction of whether the person's annual income
• Based on the attributes you provided, here is a prediction of the person's annual income:
• Based on the attributes you provided, it is likely that the person's annual income is less
• Check https://gist.github.com/danielz02/48f5672e1047dd694a66a9db1a0bb257 for more outputs.

Expected behavior

• The users should be able to change max tokens whenever needed.
• The result parsing scripts should include most of the keywords. For example, less is not included as the keyword in fairness.
• For results with a high rate of non-matching predictions, the parsing script should end with a warning or throw an error with a meaningful error message.

Proposed Fix

• Temporally increase the max_tokens settings for Llama-2-13b-chat-hf
• Add more keywords in src/dt/perspectives/fairness/score_calculation_script.py.
• Double-check previous results to see whether we had a similar problem

Environment:

• Conda Environment + Release Branch

[danielz02]

I temporarily added an option for max_tokens in the fairness configuration (c9e51f93a7029bc5568d7ac0da0414dfd75b9883). In the future, we should further refactor this argument into GenerationConfig.

[danielz02]

We also can't parse the results from the crime task properly from time to time. The examples below are from crime_br_0.0 with a similar error above.

To reproduce, use:

 dt-run +model_config=hf ++model_config.model=hf/meta-llama/Llama-2-13b-chat-hf +fairness=crime_br_0.0

[kangmintong]

• Added more key words detection for judgements
• Deal the case with a small number of expected predictions in some scenarios which possibly induce problems of fairness score calculation