A series of fine-tuned small models for an open ended version of the classic game. Premise suggested by an 8 year old.
The goal has been to have a model small enough to play the game offline or in a browser and on low end machines. Each version was fine-tuned using QLoRA via Unsloth, meaning only ~1% of the model's parameters were trained, with the rest frozen. The adapter was then merged back into the base weights and quantized to Q4_K_M GGUF format.
## Usage
**Ollama**
```bash
ollama create rockpaperanything -f Modelfile
ollama run rockpaperanything '["caterpillar", "halitosis"]'
```
```json
{"winner": "caterpillar", "loser": "halitosis", "reason": "The caterpillar's transformation from gnat food to butterfly beauty defies even the most persistent bad breath."}
```
## Input / output
Input is best done as a JSON array of two items:
```json
["arcade fire", "pie"]
```
Output is JSON:
```json
{
"winner": "arcade fire",
"loser": "pie",
"reason": "The Arcade Fire's infectious energy fills the entire venue, making even a pie feel like it needs to dance."
}
```
## Models
- v1: from Llama-3.2 3B — 1,500 examples (2.02gb as gguf)
- v2: from Llama-3.2 3B — 2,200 examples (2.02gb as gguf)
- v3: from Qwen3.5 2B — 2,200 examples (1.27gb as gguf; 1.08gb as mlc)
- v4: from Qwen3 1.7B — 2,200 examples (1.11gb as gguf; 0.98gb as mlc))
