Text Reconstruction on BBC News (test)
99.67BERTScore F1Llama-3.2-3B-Instruct (SFT)
Evaluation Results
| Method | Links | |
|---|---|---|
| Llama-3.2-3B-Instruct (SFT)rkeep=0.9, protocol=QLoRA fine-tuned (SFT), skeleton_type=WordFreq2026.05 | 99.67 | |
| Llama-3.2-3B-Instruct (ZS)rkeep=0.9, protocol=Zero-shot (ZS) with polish, skeleton_type=WordFreq2026.05 | 98.57 | |
| Gemini 2.0 Flashrkeep=0.9, protocol=zero-shot, skeleton_type=WordFreq2026.05 | 98.39 | |
| Gemini 2.0 Flashrkeep=0.7, protocol=zero-shot, skeleton_type=WordFreq2026.05 | 96.45 | |
| Llama-3.2-3B-Instruct (SFT)rkeep=0.7, protocol=QLoRA fine-tuned (SFT), skeleton_type=WordFreq2026.05 | 96.08 | |
| Llama-3.2-3B-Instruct (SFT)rkeep=0.5, protocol=QLoRA fine-tuned (SFT), skeleton_type=WordFreq2026.05 | 95.98 | |
| Llama-3.2-3B-Instruct (ZS)rkeep=0.7, protocol=Zero-shot (ZS) with polish, skeleton_type=WordFreq2026.05 | 94.65 | |
| Gemini 2.0 Flashrkeep=0.5, protocol=zero-shot, skeleton_type=WordFreq2026.05 | 93.14 | |
| Llama-3.2-3B-Instruct (SFT)rkeep=0.3, protocol=QLoRA fine-tuned (SFT), skeleton_type=WordFreq2026.05 | 92.06 | |
| Llama-3.2-3B-Instruct (ZS)rkeep=0.5, protocol=Zero-shot (ZS) with polish, skeleton_type=WordFreq2026.05 | 90.32 | |
| Gemini 2.0 Flashrkeep=0.3, protocol=zero-shot, skeleton_type=WordFreq2026.05 | 89.48 | |
| Llama-3.2-3B-Instruct (SFT)rkeep=0.1, protocol=QLoRA fine-tuned (SFT), skeleton_type=WordFreq2026.05 | 88.27 | |
| Llama-3.2-3B-Instruct (ZS)rkeep=0.3, protocol=Zero-shot (ZS) with polish, skeleton_type=WordFreq2026.05 | 87.11 | |
| Gemini 2.0 Flashrkeep=0.1, protocol=zero-shot, skeleton_type=WordFreq2026.05 | 85.27 | |
| Llama-3.2-3B-Instruct (ZS)rkeep=0.1, protocol=Zero-shot (ZS) with polish, skeleton_type=WordFreq2026.05 | 84.36 |