yuxixia/raw_results

Raw Response CSV Files

This folder contains raw model responses and confidence-estimation outputs used to evaluate confidence robustness, answer stability, and answer sensitivity.

File Naming

Most response files follow this pattern:

{dataset}_{model}_{split}_{experiment}_{prompt}[_{sample_idx}].csv

Examples:

nq_qwen72b_test_greedy_ans0.csv
popqa_olmo_train_greedy_ans0.csv
triviaqa_llama70b_test_sample_ans0_3.csv

Fields:

• dataset: one of nq, sciq, triviaqa, popqa.
• model: model key, for example mistral, llama70b, qwen72b, qwen32b, qwen14b, olmo32b, olmo13b, olmo, ministral8b, ministral14b.
• split: train or test.
• experiment:
  • greedy: greedy decoding output.
  • sample: sampled decoding output.
• prompt: answer prompt name, usually ans0 through ans9.
• sample_idx: sampled-answer index, usually 0 through 9, only present for sample files.

Files ending in _with_gpt_score.csv contain GPT-4o correctness judgments copied or generated for the corresponding response file.

Greedy Files

Greedy files contain one model answer per question under one answer prompt:

{dataset}_{model}_test_greedy_ans0.csv
...
{dataset}_{model}_test_greedy_ans9.csv

For test splits, there are usually 10 greedy files per model and dataset, one for each prompt perturbation. These are used to measure prompt robustness.

For train splits, only:

{dataset}_{model}_train_greedy_ans0.csv

is used for training calibrators.

Rows are intended to align by row order across prompt files for the same {dataset, model, split}.

Sample Files

Sample files contain stochastic answers generated with temperature sampling:

{dataset}_{model}_test_sample_ans0_0.csv
...
{dataset}_{model}_test_sample_ans0_9.csv

Each file stores one sampled answer per question. The 10 sample files for the same {dataset, model} are used to measure answer stability and sensitivity.

Rows are intended to align by row order across sample files.

Group Files

Semantic grouping files summarize equivalence classes among generated answers.

Greedy group file:

{dataset}_{model}_test_greedy_groups.csv

This groups the 10 greedy prompt answers for each question. It is used to compute prompt robustness over the semantic group containing the ans0 answer.

Sampled group file:

{dataset}_{model}_test_sampled_groups.csv

This groups the 10 sampled answers for each question. It is used to compute answer stability within the largest semantic group and answer sensitivity between semantic groups.

Important group columns:

• idx: row index used for alignment.
• question: input question.
• target: reference answer.
• semantic_groups_raw: raw GPT grouping output.
• semantic_groups: grouped answer strings in JSON format.
• semantic_group_indices: grouped answer indices in JSON format.
• greedy_answers or answers: generated answers in JSON format.
• prompt_names: prompt names for greedy grouping.

Common Response Columns

Most response CSV files contain:

• idx: row/sample index when available.
• question: original question.
• target: reference answer.
• response: generated model answer.
• response_prob: length-normalized generated sequence probability.
• boostedprob_score: boosted-probability uncertainty score.
• perplexity: perplexity computed from output logits.
• attention_score: attention-based score.
• hidden_score: hidden-state score.
• norm_attention_score: calibrated attention score.
• norm_hidden_score: calibrated hidden-state score.
• ps_score: Platt-scaled confidence score.
• p_true_score: P(True) score.
• infosel_score: Calib1 confidence score.
• vc_ans_score: verbalized confidence under answer-prompt perturbation.
• vc_conf_score: verbalized confidence under confidence-prompt perturbation.
• vc_score: verbalized confidence for sampled answers.
• saplma_score: SAPLMA probe score.
• pik_score: P(IK) probe score.
• gpt_score: GPT-4o correctness judgment, usually in _with_gpt_score.csv files.

Not every file contains every score column. Missing columns mean that method was not run for that file.

Metric Columns Saved in Group Files

After running metric evaluation, group files may also contain per-example metric columns:

• {method_prefix}_robustness: prompt robustness score, saved in greedy group files.
• {method_prefix}_stability: answer stability score, saved in sampled group files.
• {method_prefix}_sensitivity: answer sensitivity score, saved in sampled group files.

These scores use the convention that higher is better:

• Prompt robustness (P-RB) is 1 - std(confidence) over prompt answers in the semantic group containing ans0.
• Answer stability (A-STB) is 1 - std(confidence) over sampled answers in the largest semantic group.
• Answer sensitivity (A-SST) measures confidence separation between the largest and smallest semantic groups.

Example method prefixes:

• seq_likelihood
• boosted_prob
• platt_scaling
• p_true
• calib1
• verbalized_conf
• hidden_score
• attention_score
• saplma
• p_ik

Row Alignment

For downstream analysis:

• Greedy files with different prompts should be loaded together by row order for the same {dataset, model, split}.
• Sample files with different sample_idx should be loaded together by row order for the same {dataset, model, split, prompt}.
• Group files use the same row order as the corresponding response files.
• _with_gpt_score.csv files align row-by-row with their corresponding normal response CSV.

Recommended Usage

Use normal response files for confidence scores and _with_gpt_score.csv files for correctness labels. For example:

• Load nq_qwen72b_test_greedy_ans0.csv for confidence scores.
• Load nq_qwen72b_test_greedy_ans0_with_gpt_score.csv for gpt_score.
• Join them by row order.

The main evaluation metrics are:

• ECE: expected calibration error, using greedy ans0 and gpt_score.
• Brier: Brier score, using greedy ans0 and gpt_score.
• AUROC: AUROC, using greedy ans0 and gpt_score.
• P-RB: prompt robustness over greedy prompt perturbations.
• A-STB: answer stability over sampled answers in the largest semantic group.
• A-SST: answer sensitivity across sampled answer semantic groups.