Content Safety and Response Quality in LLMs: A Data-Centric Refinement Approach
Bibliographic record
Abstract
Recent advancements in large language models (LLMs) have significantly impacted natural language processing. However, ensuring the safety and quality of responses generated by LLMs remains a challenge. Building on previous work with a Corrective-BART Model, which demonstrated significant reductions in toxicity, this paper addresses the trade-of between safety and response quality. A data-centric refinement paradigm is introduced, proactively generating high-quality, safe responses during training. A dynamic dataset is curated using Llama2, where toxic prompt-response pairs are contextually regenerated into safe, relevant alternatives. The enhanced Corrective-BART model employs a multi-threshold correction pipeline, leveraging multiple metrics to detect implicit and explicit harms. For the Type-Token Ratio, the enhanced model paired with GPT-4 achieves an 11% improvement. Similarly, improvements of 8.2% and 8.6% are observed when paired with Gemma-2b-it and Mistral-7B, respectively. In terms of Readability Score, the enhanced BART model paired with GPT-4 shows an 8% improvement while demonstrating a 21% improvement when paired with Mistral-7B and an 8.4% improvement with Gemma-2b-it. For Coherence Score, the enhanced BART model achieves a 6% improvement when paired with GPT-4, a 7% improvement with Mistral-7B, and a 12.7% improvement with Gemma-2b-it. Notably, the Refusal Rate exhibits a 23.4% improvement when the enhanced BART model is paired with GPT-4. Furthermore, impressive increases of 10.5× and 21.3× are observed when paired with Mistral-7B and Gemma-2b-it, respectively. These results demonstrate that the enhanced BART consistently enhances performance across all LLMs, improving lexical diversity, readability, coherence, and safety.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame machine prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. The Gemma side is a direct model label for every work in the frame, read from the title-only record. The Codex side is a classifier learned from the 10,348 direct Codex labels and calibrated to design-weighted sample rates; fields without enough sample support carry no Codex call. Candidate is the union of the two sides; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels.
Distilled classifier scores by category (both heads)
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.008 | 0.027 |
| Meta-epidemiology (narrow) | 0.002 | 0.001 |
| Meta-epidemiology (broad) | 0.001 | 0.002 |
| Bibliometrics | 0.002 | 0.001 |
| Science and technology studies | 0.001 | 0.001 |
| Scholarly communication | 0.002 | 0.005 |
| Open science | 0.003 | 0.003 |
| Research integrity | 0.002 | 0.003 |
| Insufficient payload (model declined to judge) | 0.003 | 0.002 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one source (direct Gemma or distilled Codex), not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".