Show HN: Fine-tuned Qwen2.5-7B on 100 films for probabilistic story graphs

Source: Hacker News

Overview

Hi HN, I’m a computer systems engineering student in Mexico who switched from film school. I built CineGraphs because my filmmaker friends and I kept hitting the same wall—we had a vague idea for a film but no structured way to explore where it could go. Existing AI writing tools produced generic, formulaic output, and I didn’t want to build another ChatGPT wrapper, so I took a different route.

The idea is simple: you input a rough concept, and the tool generates branching narrative paths visualized as a graph. You can sculpt those branches into a structured screenplay format and export to Fountain for use in professional screenwriting software.

Training Data

Most AI writing tools are trained on generic internet text, which is why they output generic results. I wanted something that understood actual cinematic storytelling—not plot summaries or Wikipedia synopses, but the structural DNA of films.

• Curated 100 high‑quality films with distinctive narrative structures (e.g., Godard’s jump cuts, Kurosawa’s parallel character arcs, Brakhage’s non‑linear visual poetry, Tarkovsky’s slow‑burn temporal structures).
• Built a 1000+ line Python pipeline using Qwen3‑VL to analyze each film with subtitles enabled.
• Extracted scene‑level narrative beats, character relationships, thematic threads, and dialogue patterns.
• Iterated extensively on prompts to get the model to identify functional elements such as “this scene mirrors the opening” or “this character’s arc inverts the protagonist’s.”

From these extractions I generated a 10 K example dataset of prompt‑to‑branching‑narrative pairs.

Model Fine‑tuning

• Fine‑tuned Qwen2.5‑7B‑Instruct with a LoRA optimized for probabilistic story branching.
• The LoRA handles graph generation (exploring possible narrative directions) while the full 7B model generates the technical screenplay format on export.
• Chose the 7B model for affordable inference at scale; the entire system runs on a single RTX 4090 GPU using vLLM.

Implementation

• Frontend: React Flow for graph visualization.
• Key insight: Screenwriting is fundamentally about making choices—what if the character goes left instead of right?—but most tools force a linear document too early. The graph structure lets writers explore multiple paths before committing, matching early‑development thinking.

Results

• The film selection mattered dramatically. Early versions trained on mainstream films produced formulaic outputs. Adding experimental and international cinema greatly improved variety and interestingness.
• The model appears to treat narrative structure as a design space rather than a fixed formula.
• Used internally to overcome second‑act problems: the branching format forces consideration of possibilities rather than premature commitment.

Try It

You can test the tool at https://cinegraphs.ai/ – no signup required for a full project with up to 50 branches. Creating an account is needed only to save work; registered users receive 3 free projects.

I’d love feedback on whether the generation quality feels meaningfully different from generic AI tools, and whether the graph interface adds value or just friction.