Instruct-StoryMix Workspace

Controllable story generation experiments using component decomposition on NarrativeQA summaries. We tested whether component-conditioned prompting improves controllability versus prompt-only baselines using a small open-source model.

Key Findings

• Component-conditioned prompts reduced event/character adherence vs. prompt-only baseline (TinyLlama-1.1B, n=20).
• Plan-story similarity dropped significantly under conditioning (p≈0.034), indicating drift from provided controls.
• Style cues rarely surfaced; small model showed weak controllability.

Reproduction

1. uv venv
2. source .venv/bin/activate
3. Dependencies already in pyproject.toml (installed via uv add).
4. Run experiment: python scripts/storymix_experiment.py (uses NarrativeQA subset at datasets/narrativeqa/subset).

Outputs go to results/outputs/ (jsonl, csv) and results/metrics.json; plot at results/plots/metric_bar.png.

File Structure

• planning.md – research plan
• scripts/storymix_experiment.py – decomposition, generation, evaluation pipeline
• results/ – metrics, plots, and generated stories
• datasets/ – pre-downloaded NarrativeQA subset (see datasets/README.md)
• REPORT.md – full report