unfavalen/riv-v2

RIV v2

RIV v2 is an experimental decision-ranking model developed by Robi Labs.

Architecture

• Encoder: jhu-clsp/mmBERT-base
• Pooling: masked mean pooling
• Decision head: Linear → GELU → Dropout → Linear
• Fine-tuning: final two encoder layers unfrozen
• Maximum sequence length: 256
• Training epochs: 3
• Random seed: 42

The model scores candidate decisions using a custom neural-network head. This is not a standalone causal language model.

Evaluation

Metric	Result
Validation accuracy	59.17%
Validation examples	600
Official test accuracy	56.70%
Official test examples	2,000
Official test loss	0.9292
Expected random-choice accuracy on test set	31.75%

These are initial benchmark results from one training run. They should not be interpreted as evidence of generalization to arbitrary tasks or datasets.

Training configuration

• Encoder learning rate: 1e-5
• Decision-head learning rate: 2e-4
• Batch size: 8
• Epochs: 3
• Maximum sequence length: 256
• Seed: 42

Files

• riv_v2_checkpoint.pt: trained checkpoint
• riv_v2_metrics.json: training and evaluation metrics
• modeling_riv_v2.py: custom model architecture
• train_v2.py: training implementation
• modal_app_v2.py: Modal training entry point
• tokenizer.json and tokenizer_config.json: tokenizer files

Usage and limitations

The checkpoint uses a custom architecture and is not automatically loadable as a standard Hugging Face AutoModel without the corresponding model implementation and checkpoint-loading code.

This is an experimental research release, not a production-ready model.