modernnews-event-stream

ModernNews Event Stream

ModernNews Event Stream is a 120.3M-parameter research model that forecasts the next structured event type and a hazard score from a temporal window of news-derived events. It uses learned event, topic, and location embeddings followed by a causal TCN, an RWKV-style recurrent state, and a graph-aware prediction head.

This is not a text-generation model and does not accept raw articles. Input documents must first be normalized into the ModernNews event schema and converted into temporal tensors.

Checkpoint

This release contains the best validation checkpoint from the phase2_infini run. The checkpoint was selected at epoch 4 by lowest combined validation loss.

The hazard MAE compares the sigmoid output with the thresholded hazard target (target_hazard > 0.5). Metrics are specific to the held-out windows used by this training run and should not be treated as broad real-world benchmarks.

Architecture and input contract

• Hidden width: 768
• TCN: 8 causal residual blocks, kernel size 5
• Context window: up to 48 structured events during dataset construction (the final event is used as the target)
• Vocabulary sizes: 13 event types, 12,034 topics, and 79,186 locations
• Numeric inputs per event: time delta, severity, novelty, uncertainty, and 15 rolling features
• Outputs: next-event logits and one hazard logit

vocabs.json contains the exact training vocabularies. Unknown values should map to the <unk> entry and padding to <pad>.

Loading

from load_model import load_model

model, vocabs, config = load_model("sirus/modernnews-event-stream")

The model forward pass expects these tensors:

outputs = model(
    event_types=event_types,                    # [batch, sequence]
    attention_mask=attention_mask,              # [batch, sequence]
    time_deltas=time_deltas,                    # [batch, sequence]
    severities=severities,                      # [batch, sequence]
    novelties=novelties,                        # [batch, sequence]
    uncertainties=uncertainties,                # [batch, sequence]
    rolling_features=rolling_features,          # [batch, sequence, 15]
    topic_token_ids=topic_token_ids,            # [batch, sequence, topics/event]
    topic_token_mask=topic_token_mask,          # same shape as topic_token_ids
    location_token_ids=location_token_ids,      # [batch, sequence, locations/event]
    location_token_mask=location_token_mask,    # same shape as location_token_ids
    adjacency=adjacency,                        # optional [batch, batch]
)

next_event_ids = outputs.event_logits.argmax(dim=-1)
hazard_probabilities = outputs.hazard_logits.sigmoid()

The bundled implementation uses a portable PyTorch recurrent fallback. It may optionally compile the included CPU extension for faster RWKV-style scanning when a compiler toolchain is available. Set MODERNNEWS_DISABLE_RWKV_EXT=1 to force the fallback.

License

The model weights (event_model.pt) and vocabularies (vocabs.json) are licensed under the Creative Commons Attribution-NonCommercial 4.0 International License. You may share and adapt those artifacts with attribution, but you may not use them for commercial purposes. Contact the model author through Hugging Face to discuss separate commercial licensing.

The bundled implementation source (load_model.py and modernnews/) remains available under the MIT License. See LICENSE for the exact file-level split.

Training data

The phase-2 pipeline combines structured and news-derived event streams, including automatically normalized or teacher-labeled records. Some labels are synthetic or weakly supervised. The release does not include the training corpus.

Limitations and intended use

This checkpoint is a research prototype for experimentation with aggregate event streams. It can inherit geographic, linguistic, source-selection, and labeling biases from its source data. Predictions may be poorly calibrated under distribution shift, and event streams can contain noisy, duplicated, delayed, or incorrect records.

Do not use this model as the sole basis for emergency response, public safety, military, intelligence, legal, financial, employment, insurance, medical, or other high-impact decisions. Human review and independent corroboration are required for consequential use.