cs.CLApr 13, 2026
SaveBITS Pilani at SemEval-2026 Task 9: Structured Supervised Fine-Tuning with DPO Refinement for Polarization Detection
Organizations: Birla Institute of Technology and Science, Pilani, India
Abstract
The POLAR SemEval-2026 Shared Task aims to detect online polarization and focuses on the classification and identification of multilingual, multicultural, and multi-event polarization. Accurate computational detection of online polarization is challenging due to nuanced rhetoric, implicit framing, and the high cost of human-in-the-loop annotation. Building on recent findings that contextual prompting enables large language models to function as strong polarization detectors, we present a two-stage approach for detecting polarization in social media text that combines structured supervised fine tuning with Direct Preference Optimization (DPO) refinement. We fine tune Qwen 2.5-7B-Instruct with LoRA using an interpretable slot-filling template (target, claim type, manifestation checklist, and justification). We then apply DPO with automatically generated preference pairs to reduce costly false negatives. Our submitted system achieves 0.7664 Macro-F1 on the English test set. Post-submission experiments with Mistral-Nemo-Instruct-2407 and LLM-judge-filtered preference pairs further improve to 0.8162 Macro-F1 (not submitted to CodaBench), surpassing the organiser baseline of 0.7802.
Figures & tables
Figure 1: Overview of our two-stage pipeline for Subtask 1 polarization detection: structured supervised fine-tuning (SFT) first generates schema-consistent outputs, and DPO refinement then learns from preference pairs to produce the final prediction.
| Split | Total | Polarized | Non-polarized |
|---|---|---|---|
| Train | 3,222 | 1,175 (36.5%) | 2,047 (63.5%) |
| Dev | 160 | 59 (36.9%) | 101 (63.1%) |
| Test | 1,452 | 533 (36.7%) | 919 (63.3%) |
| Total | 4,834 | 1,767 (36.6%) | 3,067 (63.4%) |
Table 1: English dataset statistics and label distribution across official splits.
| Method | English Dev F1 |
|---|---|
| Zero-shot baseline | 0.7105 |
| SLMs | 0.7149 |
| SFT | 0.738 |
| SFT + DPO | 0.7893 |
Table 2: F1 scores on the English development set. SLMs = Small Language Models baseline (DistilBERT fine-tuned for sequence classification). Weighted loss during SFT did not improve performance.
| System | Rank (out of 60) | F1 |
|---|---|---|
| Highest-ranked system | 1 | 0.8252 |
| POLAR baseline | 47 | 0.7802 |
| Our system | 52 | 0.7664 |
Table 3: Unofficial English Subtask 1 leaderboard comparison (60 teams).
| Metric | SFT | SFT + DPO |
|---|---|---|
| Accuracy | 0.7812 | 0.8000 |
| Precision | 0.8333 | 0.7077 |
| Recall | 0.5085 | 0.7797 |
| F1 (Binary) | 0.6316 | 0.7419 |
| F1 (Macro) | 0.7380 | 0.7893 |
| F1 (Micro) | 0.7812 | 0.8000 |
Table 4: English development set metrics comparing SFT vs. SFT + DPO.
| Model | mF1 | Acc. | P(1) | R(1) |
|---|---|---|---|---|
| Mistral-Nemo | 0.7963 | 0.8030 | 0.7928 | 0.8121 |
| Qwen2.5-7B | 0.7835 | 0.7899 | 0.7811 | 0.8006 |
Table 5: Label-only SFT performance on the English test set ( ) trained on the original unmodified labels without the Rejudged Sonnet dataset.
| Mode | F1 | FNs | FPs |
|---|---|---|---|
| SFT | 0.7795 | 158 | 137 |
| SFT with DPO (330 pairs) | 0.7889 | 132 | 155 |
| SFT with DPO (721 pairs) | 0.7637 | 64 | 274 |
Table 6: Effect of DPO preference pair count on F1, false negatives (FNs), and false positives (FPs). SFT baseline is the 10-epoch full-training-set model; development-set results appear in Tables 2 and 4 .
| Condition | Acc. | P(1) | R(1) | mF1 |
|---|---|---|---|---|
| Label-only SFT | 0.792 | 0.777 | 0.7884 | 0.781 |
| Label-only + DPO | 0.720 | 0.618 | 0.625 | 0.699 |
| Reasoning SFT | 0.793 | 0.745 | 0.662 | 0.771 |
| Reasoning + DPO | 0.802 | 0.732 | 0.704 | 0.789 |
Table 7: Structured rationale ablation on the English test set. P(1) and R(1) = precision and recall for the polarized class; mF1 = Macro-F1. Best in bold.
| System | mF1 |
|---|---|
| Qwen2.5-7B Label-Only SFT (Rejudged) | 0.7811 |
| Mistral-Nemo Label-Only SFT (Rejudged) | 0.8019 |
| Mistral-Nemo SFT (Rejudged) | 0.8097 |
| Mistral-Nemo DPO (FP:FN 62:38, =0.3) | 0.8162 |
Table 8: Post-submission results on the English test set ( ). None submitted to CodaBench. mF1 = Macro-F1.
Appendix figures & tables7 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 2: Post-submission pipeline. Rejudged Sonnet training data produces an SFT checkpoint, which then generates completions at varied temperatures to form candidate preference pairs. DeepSeek-R1 filters these into a balanced set (62:38 FP:FN ratio). DPO then refines the same SFT checkpoint (dashed arrow) using the filtered pairs, producing the best post-submission result of 0.8162 Macro-F1. The Rejudged Sonnet training data was produced by generating reasoning with GPT OSS 120B Nitro and filtering labels with Claude 3.5 Sonnet as a judge (see Appendix A.7 ).
| mF1 | Acc. | P(1) | R(1) | |
|---|---|---|---|---|
| 0.10 | 0.8127 | 0.8230 | 0.739 | 0.801 |
| 0.15 | 0.8125 | 0.8230 | 0.740 | 0.799 |
| 0.20 | 0.8136 | 0.8237 | 0.738 | 0.805 |
| 0.25 | 0.8108 | 0.8209 | 0.734 | 0.803 |
| 0.30 | 0.8162 | 0.8230 | 0.736 | 0.809 |
| 0.35 | 0.8079 | 0.8182 | 0.731 | 0.799 |
Table 9: Beta sweep on 299 DeepSeek-R1-filtered DPO pairs (62:38 FP:FN ratio), Mistral-Nemo-Instruct-2407, 2 epochs. mF1 = Macro-F1; P(1)/R(1) = precision/recall for the polarized class.
| Input text ( eng_03a6df98 ) |
|---|
| “Ending 2024 with an illadvised text conversation with my brother that ended with ‘Really little bro? I can deal with the atheism, but socialism?”’ |
| Gold label: 1 (Lack of Empathy: present; Invalidation: present) |
| SFT output (label: 0 — false negative) |
| Target: specific individual Claim type: evaluative opinion |
| Stereotype: absent Vilification: absent |
| Dehumanization: absent Extreme Language: absent |
Table 10: Qualitative comparison of SFT (Mistral-Nemo baseline) vs. best DPO model (299 DeepSeek-R1-filtered pairs, =0.3) on a false-negative example from the test set. Bold fields differ between the two outputs. SFT misreads the implicit ideological framing as a personal anecdote and predicts 0. DPO recovers the correct label (1) by identifying Lack of Empathy and Invalidation, though it also flags Stereotype and Vilification as present, which are absent in the gold annotation.
| Parameter | Value |
|---|---|
| Max Length | 1024 |
| Train Batch Size | 1 |
| Eval Batch Size | 4 |
| Gradient Accumulation Steps | 8 |
| Learning Rate | 5e-5 |
| Num Epochs | 3 |
Table 11: Pre-submission SFT hyperparameters (Qwen2.5-7B-Instruct). Class weights [1.0, 1.5] were used in both pre- and post-submission runs but did not massively improve performance over unweighted training (see Section 5 ).
| Parameter | Value |
|---|---|
| Max Length | 1024 |
| Max Prompt Length | 512 |
| Train Batch Size | 1 |
| Gradient Accumulation Steps | 8 |
| Learning Rate | 5e-6 |
| Num Epochs | 2 |
Table 12: DPO hyperparameters. Pre-submission uses Qwen2.5-7B-Instruct with =0.1; post-submission uses Mistral-Nemo-Instruct-2407 with =0.3.
| Configuration | FP:FN | Total | mF1 | Recall |
|---|---|---|---|---|
| SFT baseline | — | — | 0.8097 | 0.827 |
| Original 190/70 | 73:27 | 260 | 0.8012 | 0.747 |
| Subset 150/70 | 68:32 | 220 | 0.8024 | 0.745 |
| Subset 120/70 | 63:37 | 190 | 0.8005 | 0.779 |
| Subset 100/70 | 59:41 | 170 | 0.7986 | 0.739 |
Table 13: DPO subset sweep on the stale unfiltered preference pairs: effect of subsampling class-0 (FP-correcting) pairs while holding class-1 (FN-correcting) pairs fixed at 70. Section 6.3 ), confirming that pair quality and volume rather than FP:FN ratio alone drive the performance gap. Matching the test-set class distribution (120/70, 63:37) did not improve performance.
| Parameter | Value |
|---|---|
| Base Model | Qwen/Qwen2.5-7B-Instruct |
| Training Data | 3,222 samples (original) |
| Max Length | 1024 |
| Batch Size | 8 |
| Gradient Accumulation | 1 |
| Learning Rate | 5e-5 |
Table 14: Label-only SFT hyperparameters (Qwen2.5-7B-Instruct).