| --- |
| language: en |
| license: mit |
| base_model: distilroberta-base |
| pipeline_tag: text-classification |
| tags: |
| - news |
| - sports |
| - movies |
| - multi-task |
| --- |
| |
| # MAPNAI news classifier (Agent 2) |
|
|
| Multi-task news classifier used by Agent 2 of the MAPNAI pipeline. A fine-tuned |
| `distilroberta-base` encoder with four heads: |
|
|
| | Output | Values | |
| |---|---| |
| | `domain` | `entertainment_movies`, `sports`, `other` | |
| | `category` | Box Office, Reviews, Releases & Trailers, Casting & Production, Awards & Festivals, Celebrity News, Streaming/OTT 路 Cricket, Football, Tennis, Basketball, Motorsport, Olympics & Other Sports 路 Other | |
| | `urgency_flag` | true / false (threshold stored in `label_config.json`) | |
| | `sentiment` | -1.0 to 1.0 | |
|
|
| This is a custom architecture, so it is not loadable with `AutoModelForSequenceClassification`. |
| Load it with `agents/classifier_model.py` from the MAPNAI repository: |
|
|
| ```bash |
| hf download satvik4577/mapnai-classifier --local-dir models/classifier |
| ``` |
|
|
| ```python |
| from agents.classifier_model import NewsClassifierPredictor |
| p = NewsClassifierPredictor("models/classifier") |
| p.predict(["Kohli ruled out of the World Cup final with a hamstring injury"]) |
| ``` |
|
|
| **Training data:** news from public RSS feeds plus the AG News, BBC News and HuffPost News |
| Category datasets, labelled by Qwen2.5-7B-Instruct and validated against the taxonomy. |
|
|
| **Version 1 notes:** movie categories lean towards "Box Office"; retraining with class weighting |
| and targeted data is planned. |
|
|
