| # Generated data | |
| Training sets we generated or augmented for the released models. Each row is `text, target, stance` | |
| with `stance` in `{Favor, Against, None}`. Paths match what the configs and training scripts expect, | |
| so run them with this folder as the data root. | |
| | File | Rows | How it was made | Used by | | |
| |---|---|---|---| | |
| | `track2/train_aug.csv` | 8,056 | Track-2 train plus LLM paraphrases; a `source` column marks `original` (2,714) or `paraphrase` (5,342) | encoder `t2_marbert_aug`, via `configs/track2_aug.yaml` | | |
| | `synth/t1_train_style.csv` | 3,902 | Track-1 train plus style-grounded generated shots | LoRA `allam_style` | | |
| | `external/pool_real_all.csv` | 4,449 | pooled real-target shots (Women Empowerment, Covid, Digital) | LoRA `allam_real` | | |
| The official Mawqif-v2 splits belong to the organizers and are not redistributed here. Get them from | |
| the shared task at https://stanceeval.github.io/. Everything above is derived from those splits and | |
| can be regenerated: | |
| ```bash | |
| python -m src.gen_synth --help # style and target-matched shot pools | |
| python -m src.augment --help # paraphrase augmentation | |
| ``` | |