Automated Essay Scoring on IELTS Writing Task 2
99.02AccuracyLlama-3.1-8B (4-LoRA)
Evaluation results
| Method | Links | ||||
|---|---|---|---|---|---|
| Llama-3.1-8B (4-LoRA)Approach=A3 + RAG, Training / Prompting Scheme=k-Instruction Tuning + RAG, k-shot=2-shot2026.03 | 99.02 | 93.5 | 0.87 | 0.62 | |
| Llama-3.1-8B (1-LoRA)Approach=A4, Training / Prompting Scheme=SFT + DPO + RAG, k-shot=2-shot2026.03 | 98.7 | 92.5 | 0.84 | 0.58 | |
| Llama-3.1-8B (4-LoRA)Approach=A3 + RAG, Training / Prompting Scheme=k-Instruction Tuning (w/o RAG), k-shot=2-shot2026.03 | 98.18 | 79.99 | 0.9947 | 0.7676 | |
| Llama-3.1-8B (1-LoRA)Approach=A3 + RAG, Training / Prompting Scheme=k-Instruction Tuning + RAG, k-shot=2-shot2026.03 | 97.5 | 88 | 0.92 | 0.66 | |
| Llama-3.1-8B (not finetuned)Approach=A3 + RAG, Training / Prompting Scheme=k-Instruction Tuning + RAG, k-shot=2-shot2026.03 | 96.3 | 84 | 0.98 | 0.7 | |
| Gemini 2.5 ProApproach=A2, Training / Prompting Scheme=Prompting Tuning, k-shot=zero-shot2026.03 | 96.16 | 55.81 | 1.4977 | 1.2026 | |
| Mistral-7B-Instruct-v0.3 (1-LoRA)Approach=A4, Training / Prompting Scheme=SFT + DPO + RAG, k-shot=2-shot2026.03 | 94.7 | 87.5 | 1.03 | 0.83 | |
| Mistral-7B-Instruct-v0.3 (not finetuned)Approach=A3 + RAG, Training / Prompting Scheme=k-Instruction Tuning + RAG, k-shot=2-shot2026.03 | 90.5 | 72 | 1.18 | 0.85 | |
| Llama-3.1-8BApproach=A3, Training / Prompting Scheme=Instruction Tuning + Few-shot, k-shot=4-shot2026.03 | 87.08 | 56.79 | 1.3547 | 0.9436 | |
| Llama-3.1-8BApproach=A3, Training / Prompting Scheme=Instruction Tuning + Few-shot, k-shot=2-shot2026.03 | 86.8 | 52 | 0.94 | 1.02 | |
| Phi-2Approach=A3, Training / Prompting Scheme=Instruction Tuning, k-shot=zero-shot2026.03 | 78.12 | 41 | 1.35 | 0.6505 | |
| Llama-3.1-8BApproach=A3, Training / Prompting Scheme=Instruction Tuning, k-shot=zero-shot2026.03 | 75.96 | 45.11 | 1.0172 | 0.9859 | |
| GPT-2 (LLM-based)Approach=A1, Training / Prompting Scheme=All Parameters, k-shot=–2026.03 | 73.5 | 30.9 | 0.77 | 0.539 | |
| RoBERTa-base (BERT-based)Approach=A1, Training / Prompting Scheme=All Parameters, k-shot=–2026.03 | 73.1 | 30.4 | 0.784 | 0.549 | |
| GPT-4oApproach=A2, Training / Prompting Scheme=Prompting Tuning, k-shot=zero-shot2026.03 | 72 | 29.2 | 1.13 | 0.791 | |
| GPT-4oApproach=A2, Training / Prompting Scheme=Prompting Tuning, k-shot=2-shot2026.03 | 71 | 28.1 | 1.05 | 0.735 | |
| GPT-2 (LLM-based)Approach=A1, Training / Prompting Scheme=Classifier Only, k-shot=–2026.03 | 70.7 | 27.8 | 0.757 | 0.53 | |
| Mistral-7BApproach=A3, Training / Prompting Scheme=Instruction Tuning, k-shot=zero-shot2026.03 | 70 | 27 | 1.4128 | 1.1214 | |
| Gemma-7BApproach=A3, Training / Prompting Scheme=Instruction Tuning, k-shot=zero-shot2026.03 | 67 | 22 | 1.33 | 0.98 | |
| RoBERTa-base (BERT-based)Approach=A1, Training / Prompting Scheme=Classifier Only, k-shot=–2026.03 | 65.1 | 21.6 | 0.83 | 0.581 | |
| Llama-3-70BApproach=A2, Training / Prompting Scheme=Prompting Tuning, k-shot=zero-shot2026.03 | 63 | 19.3 | 0.99 | 0.693 | |
| Llama-3-70BApproach=A2, Training / Prompting Scheme=Prompting Tuning, k-shot=2-shot2026.03 | 56 | 11.6 | 1.25 | 0.875 |