Research map / Large language models / Instruction and fine-tuning
Enhancing Trustworthiness of Fine-Tuned LLMs via Regularized Subset Selection
Kumar Shubham, Nishant Sharma, Karn Tiwari, Prathosh AP
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 |
|---|---|---|---|
| apply PBRF repair on selected detrimental subset after SFT | Qwen2.5-7B post-SFT · Truthfulness | log-odds: 0.732 -> 0.662 | Improved |
| apply PBRF repair on selected detrimental subset after SFT | Pythia-6.9B post-SFT · Stereotypical bias | log-odds: −0.380 -> −0.449 | Improved |
| apply PBRF repair after SFT | Qwen2.5-7B post-SFT · Machine ethics | log-odds: −0.253 -> −0.275 | Improved |
| apply PBRF repair after SFT | post-SFT models · Anthropic HH perplexity | perplexity: 5.401 -> 5.419 | No clear effect |
| repair with a common subset for all trust metrics instead of metric-specific subsets | SeerRepair common subset · Truthfulness | relative log-odds change %: +21.73 | Mixed |
| add PBRF-based update instead of GA/SGA/GA+KL gradient ascent (ablation) | trust repair · trust metrics | log-odds and perplexity | Improved |
| select subset with DPP/attribution instead of random subset (ablation) | gradient ascent repair · Machine ethics, Pythia-6.9B | perplexity relative change %: -26.2 | Improved |
| add DPP diversity regularization to subset selection (ablation) | repair method, Pythia-1.4B and Qwen2.5-1.5B · higher learning rates | perplexity | Improved |
| use PBRF repair instead of retraining without detrimental samples | Pythia-1.4B · Truthfulness | log-odds: 0.504 -> 0.476 | No clear effect |
| use repair instead of DPO | Qwen2.5-7B · Truthfulness | log-odds: 0.787 -> 0.573 | Improved |
| use repair instead of BiPO steering | Pythia-1.4B · Stereotypical bias | relative log-odds change %: 4.339 -> 13.430 | Improved |
| apply repair to Qwen2.5 models | Qwen2.5-7B · GSM8K accuracy | accuracy: 80.89 -> 80.14 | No clear effect |
Closest papers by meaning
- From Yes-Men to Truth-Tellers: Addressing Sycophancy in Large Language Models with Pinpoint Tuning (ICML 2024 · similarity 0.90)
- Position: TrustLLM: Trustworthiness in Large Language Models (ICML 2024 · similarity 0.90)
- Token Cleaning: Fine-Grained Data Selection for LLM Supervised Fine-Tuning (ICML 2025 · similarity 0.89)
- Enhancing Multiple Dimensions of Trustworthiness in LLMs via Sparse Activation Control (NeurIPS 2024 · similarity 0.89)
- Data Mixing Optimization for Supervised Fine-Tuning of Large Language Models (ICML 2025 · similarity 0.88)
- HonestLLM: Toward an Honest and Helpful Large Language Model (NeurIPS 2024 · similarity 0.88)
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.
