Seinfeld Scene Generator — AI dialogue in your browser

Generate a Seinfeld scene

TOPIC / SITUATION

How it works

Two language models were fine-tuned on ~0 Seinfeld script excerpts to learn the show's dialogue style, character voices, and scene structure. Both models run entirely in your browser using WebAssembly — no server, no API keys, no data leaves your machine.

Llama 3.2 3B was trained with QLoRA (quantized low-rank adaptation) — only 0 of the model's parameters were updated, but that was enough to teach it Seinfeld's format and tone. The model is served as a 4-bit GGUF file (~0, sharded into 4 chunks for browser memory limits) and executed via WebAssembly with optional multi-threading.

GPT-2 Medium (0 params) was fine-tuned with a deeper LoRA configuration targeting both attention and MLP layers (r=64, 0 trainable params). It's served as an int8 ONNX model (~0) and runs through the ONNX Runtime WebAssembly backend.

Generation pipeline

When you click "generate", here's what happens under the hood:

The Llama backend uses multi-round generation: it generates the first character's turn, then injects the next character's name into the prompt and generates again, repeating 4 times to ensure all main characters speak. Each round uses top-k sampling (k=8, temp=0.7) with a repetition penalty.

TOPIC: losing a parking spot

CHARACTERS: JERRY, GEORGE, ELAINE, KRAMER

[JERRY'S APARTMENT]

JS-side filtering

Raw model output is noisy — small models hallucinate character names, repeat phrases, and ramble past natural endpoints. A 6-stage post-processing pipeline cleans this up entirely in JavaScript before rendering:

• 1. Character typo correction — fixes common hallucinations: JERREY→JERRY, GEROGE→GEORGE, KRAMRE→KRAMER (15 known patterns)
• 2. Punctuation normalization — collapses !!!!! to !!, strips stray brackets, ensures spaces after periods
• 3. Repetition removal — deduplicates repeated sentences and detects looping n-gram patterns
• 4. Monologue capping — trims any single character turn to 80 words max, cutting at sentence boundary
• 5. Trailing trim — cuts off incomplete sentences at the end of generation
• 6. Scene parsing — extracts [LOCATION] tags and CHARACTER: text pairs from both line-by-line and inline formats

LoRA fine-tuning

Both models were trained on the same dataset of 2,295 Seinfeld script excerpts, formatted as:

TOPIC: Jerry finds out his new girlfriend is a close talker

[JERRY'S APARTMENT]

JERRY: So I'm standing there, and she's like six inches from my face.
GEORGE: Six inches? That's nothing. I had a woman once...
...
[END]

LoRA (Low-Rank Adaptation) freezes the original model weights and inserts small trainable matrices into the attention layers. This means we only update ~1-7% of the parameters, which is enough to teach the model the Seinfeld format and character voices while keeping its general language ability intact.

The Llama model used QLoRA (4-bit quantized base + LoRA adapters) with r=32 and alpha=64, trained for 5 epochs on a single A100 GPU in ~25 minutes. GPT-2 used a deeper LoRA with r=64 targeting both attention and MLP layers, trained for 20 epochs in ~19 minutes.

We tried Qwen2.5-7B (both base and instruct) but the base model's code/math priors were too strong, and the instruct model's RLHF training fought our plain-text format. Smaller models with weaker priors turned out to be easier to steer.

Libraries