Bimaldo/riddle-diffusion-phase3

1---2language: en3license: mit4tags:5- diffusion6- text-generation7- riddles8- diffusion-lm9datasets:10- prajwaldongre/riddles-a-synthetic-riddle-dataset-for-nlp11metrics:12- exact-match13- token-f114---15 16# Diffusion-LM Riddle Solver — Phase 3 Reconstructed17 18A Diffusion-LM-style (Li et al., 2022) text generation model trained on 232 synthetic riddles. Held-out exact match: **47.0%** (K=1 and K=10).19 20## Model description21 22Continuous embedding diffusion with a Transformer encoder/decoder. The model takes a riddle as context and iteratively denoises Gaussian noise into answer word embeddings via a learned reverse process.23 24## Intended use25 26Research and diagnostics. Not a production-ready riddle solver.27 28## Training data29 30[Riddles — A Synthetic Riddle Dataset for NLP](https://www.kaggle.com/datasets/prajwaldongre/riddles-a-synthetic-riddle-dataset-for-nlp) (CC0). 232 training examples after deduplication.31 32## Architecture33 34| Parameter | Value |35|-----------|-------|36| Parameters | 8,024,576 |37| Timesteps (T) | 200 |38| d_model | 256 |39| Layers | 4 |40| d_ff | 1024 |41| Heads | 4 |42| Answer length | 4 |43| Noise schedule | sqrt power-law |44 45## Performance46 47| Split | Exact match (K=1) | Token F1 |48|-------|-------------------|----------|49| Train (n=192) | 87.5% | 0.960 |50| Held-out (n=66) | 47.0% | 0.523 |51 52## Limitations53 54- Trained on 232 examples only. Does not generalize broadly.55- Uses continuous embedding diffusion with Euclidean clamping. Discrete formulations may differ.56- Exact-match metric penalizes semantically equivalent answers.57 58## Files59 60- `model.safetensors`: Model weights61- `config.json`: Architecture hyperparameters62- `vocab.json`: Vocabulary mapping63- `inference.py`: Standalone prediction script64 65## Source66 67Full source code, diagnostics, and reproduction configs: [github.com/beme08/riddle-diffusion-lm](https://github.com/beme08/riddle-diffusion-lm)68