| language: | |
| - en | |
| pipeline_tag: text-classification | |
| tags: | |
| - political-nlp | |
| - political-ideology | |
| - viewpoint-classification | |
| - political-slant | |
| - us-politics | |
| - stance-detection | |
| - binary-classification | |
| - ensemble | |
| - mixture-model | |
| - probability-calibration | |
| - mistral | |
| - qwen | |
| - lora | |
| - mlx | |
| - apple-silicon | |
| - english-text-classification | |
| base_model: | |
| - mlx-community/Mistral-7B-v0.3-4bit | |
| - mlx-community/Qwen3.5-9B-4bit | |
| # PoSSUM | |
|  | |
| David Driscoll 路 Political Slant Scoring Using Mixtures. Classifies the viewpoint expressed in short English passages about U.S. politics as left or right. | |
| **Frozen MITweet topic test (159 passages)** 路 Macro-F1 **0.890** 路 balanced accuracy **0.872** 路 log loss **0.241**. | |
| Download with `hf download ddrisco11/PoSSUM --local-dir PoSSUM`. The bundle contains the Mistral LoRA, both calibrators and mixture weights. MLX inference requires the separately downloaded base models listed above. | |
| [GitHub and inference code](https://github.com/ddrisco11/PoSSUM) 路 [Evaluation report and pinned base revisions](https://github.com/ddrisco11/PoSSUM/blob/main/reports/human-model-release.json) | |
| **Sources** 路 [MITweet dataset and citation](https://github.com/LST1836/MITweet) 路 [Synthetic development data](https://huggingface.co/datasets/JyotiNayak/political_ideologies) 路 [Mistral 7B v0.3](https://huggingface.co/mistralai/Mistral-7B-v0.3) 路 [Qwen3.5-9B](https://huggingface.co/Qwen/Qwen3.5-9B) 路 [Hu et al., LoRA](https://arxiv.org/abs/2106.09685). | |