Research map / Architectures and efficiency / Pruning and efficient tuning
TANDA: Transfer and Adapt Pre-Trained Transformer Models for Answer Sentence Selection
Siddhant Garg, Thuy Vu, Alessandro Moschitti
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Transfer on ASNQ, then adapt on the target dataset instead of single-step fine-tuning | BERT and RoBERTa answer sentence selection models · WikiQA and TREC-QA test clean | MAP: 0.813 -> 0.893; 0.857 -> 0.912 | Improved |
| Use sequential ASNQ transfer and target adaptation instead of jointly fine-tuning on ASNQ and TREC-QA | BERT-Base · TREC-QA test clean | MAP: 0.898 -> 0.912 | Improved |
| Add target adaptation after ASNQ transfer | RoBERTa-Large-MNLI · WikiQA test clean | MAP: 0.910 -> 0.920 | Improved |
| Add target adaptation after ASNQ transfer | RoBERTa-Large-MNLI · TREC-QA test clean | MAP: 0.880 -> 0.943 | Improved |
| Use ASNQ rather than QNLI as the transfer dataset (ablation) | BERT-Base TANDA · WikiQA test clean | MAP: 0.832 -> 0.893 | Improved |
| Use ASNQ rather than QNLI as the transfer dataset (ablation) | BERT-Base TANDA · TREC-QA test clean | MAP: 0.863 -> 0.912 | Improved |
| Use only ASNQ negative label 3 rather than all negative labels (ablation) | BERT-Base fine-tuned on ASNQ · WikiQA test clean | MAP: 0.884 -> 0.881 | No clear effect |
| Use only ASNQ negative label 2 rather than all negative labels (ablation) | BERT-Base fine-tuned on ASNQ · WikiQA test clean | MAP: 0.884 -> 0.751 | Hurt |
| Use only ASNQ negative labels 2 and 3 rather than labels 1, 2, and 3 (ablation) | BERT-Base fine-tuned on ASNQ · WikiQA test clean | MAP: 0.884 -> 0.883 | No clear effect |
| Inject 20% label noise into target training data while using TANDA rather than single-step fine-tuning | BERT-Base · WikiQA test clean | MAP: 22.63% -> 2.57% drop | Improved |
| Inject 20% label noise into target training data while using TANDA rather than single-step fine-tuning | BERT-Base · TREC-QA test clean | MAP: 13.88% -> 2.30% drop | Improved |
| Use ASNQ transfer followed by NAD adaptation rather than fine-tuning only on NAD | BERT and RoBERTa models · Alexa Sample 1, Sample 2, and Sample 3 | MAP: 0.506 -> 0.585; 0.432 -> 0.474; 0.352 -> 0.391 | Improved |
Closest papers by meaning
- TAIA: Large Language Models are Out-of-Distribution Data Learners (NeurIPS 2024 · similarity 0.89)
- Get more for less: Principled Data Selection for Warming Up Fine-Tuning in LLMs (ICLR 2024 · similarity 0.85)
- CoLoR-Filter: Conditional Loss Reduction Filtering for Targeted Language Model Pre-training (NeurIPS 2024 · similarity 0.85)
- A Semantic-Sampling Framework for Evaluating Calibration in Open-Ended Question Answering (NeurIPS 2026 · similarity 0.85)
- BLADE: Enhancing Black-Box Large Language Models with Small Domain-Specific Models (AAAI 2025 · similarity 0.85)
- CABINET: Content Relevance-based Noise Reduction for Table Question Answering (ICLR 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
