Task-Specific Knowledge Distillation for Accurate and Efficient Text Summarization
Bibliographic record
Abstract
Generative AI has emerged as a transformative technology in natural language processing (NLP). It is enabling advanced capabilities such as text summarization, question answering, and content generation. Large Language Models (LLMs) have demonstrated exceptional performance in NLP tasks, with text summarization being a prominent example. Proprietary LLMs, such as GPT-4 and LLaMA 70B, achieve high accuracy but often incur substantial computational costs, usage fees, limited accessibility, and potential data privacy risks. In contrast, compact LLMs, including LLaMA 3.1 8B and Falcon 7B, offer greater flexibility, transparency, and control but frequently suffer from factual inconsistencies and semantic errors. In this study, we propose a task-specific knowledge distillation (KD) technique to transfer summarization capabilities from large teacher models LLaMA 3.1(70B), Falcon (40B), Gemma2 (27B), and Qwen $2.5(72 \mathrm{~B})$ to smaller student models $(8 \mathrm{~B}, 7 \mathrm{~B}, 2 \mathrm{~B}$, and 7 B, respectively). The distillation process leverages both cross-entropy loss and Kullback-Leibler divergence to align student predictions with teacher outputs. Distilled models are evaluated on Semantic Textual Similarity (STS-B) and Multi-Genre Natural Language Inference (MNLI) tasks, along with response time metrics. In experiments, the LLaMA 3.1 8B distilled student model achieves 0.85 STS-B Pearson correlation and 0.81 MNLI accuracy, retaining over $90 \%$ of the teacher model’s performance while reducing response time from 12s to 3s. Overall, distilled student models retain $\mathbf{8 5 - 9 0 \%}$ of the teachers’ semantic and factual performance while reducing inference latency by $\mathbf{3}-\mathbf{6} \boldsymbol{\times}$. These findings demonstrate that logit-based knowledge distillation enables the development of accurate and efficient summarization models suitable for resource-constrained environments. Entire code implementation can be found at:https://github.com/Abishethvarman/KD-Text-Summarization
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 distilled prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. Learned from the 10,348 direct Codex labels and 10,348 direct Gemma labels. Candidate is the union of thresholded teacher heads; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels or direct frontier model labels.
Codex and Gemma teacher scores by category
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.000 | 0.000 |
| Meta-epidemiology (narrow) | 0.000 | 0.000 |
| Meta-epidemiology (broad) | 0.000 | 0.000 |
| Bibliometrics | 0.000 | 0.001 |
| Science and technology studies | 0.000 | 0.000 |
| Scholarly communication | 0.001 | 0.000 |
| Open science | 0.000 | 0.000 |
| Research integrity | 0.000 | 0.000 |
| Insufficient payload (model declined to judge) | 0.000 | 0.000 |
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 teacher head, 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".