Files
Nemotron-nano-4b-escape-room/README.md
ModelHub XC d4b59a67e3 初始化项目,由ModelHub XC社区提供模型
Model: build-small-hackathon/Nemotron-nano-4b-escape-room
Source: Original Platform
2026-07-23 22:24:15 +08:00

75 lines
3.4 KiB
Markdown

---
license: other
license_name: nvidia-nemotron-open-model-license
license_link: >-
https://www.nvidia.com/en-us/agreements/enterprise-software/nvidia-nemotron-open-model-license/
tags:
- text-generation
- finetuned
- nemotron
- llama-cpp
- GGUF
language:
- en
base_model: unsloth/NVIDIA-Nemotron-3-Nano-4B-GGUF
datasets:
- build-small-hackathon/1000-Rooms-DS
---
# Fine-tuned Nemotron-3 Nano 4B
This model is a fine-tuned version of **NVIDIA Nemotron-3 Nano 4B**, adapted for improved performance on a custom dataset for generating escape rooms. This fine-tune was realised for the build-small-hackathon using **Unsloth** and using compute credits provided by **Modal**.
---
## Model Details
- **Base model:** https://huggingface.co/nvidia/Nemotron-3-Nano-4B
- **Quantized model:** https://huggingface.co/unsloth/NVIDIA-Nemotron-3-Nano-4B-GGUF
- **Architecture:** Decoder-only Transformer
- **Parameters:** ~4B
- **Fine-tuning method:** LoRA
- **Framework:** Unsloth
- **Training environment:** Modal
---
## Intended Use
This model is intended for:
- Experimental / research use
- Generating room titles and descriptions / generating responses for trying to open doors/containers as json
It has not been optimized for production safety or alignment.
---
## Limitations
- May produce incorrect or hallucinated outputs
- Not guaranteed to follow instructions perfectly
- Sensitive to prompt formatting
- May inherit biases from base model and training data
---
## How to Use
### llama.cpp
```bat
llama-cli.exe -m nemotron-room-lora-q4-k-m.gguf ^
-n 2048 --temp 0.7 --repeat-penalty 1.1 ^
--grammar-file json.gbnf ^
-sys "You are a creative dungeon-master AI that generates escape-room content. Always respond with a single valid JSON object and nothing else." ^
-p "{\"task\":\"generate_room\"} Respond with room_name, room_story, room_prompt, door_description, door_prompt, door_key_name, door_key_prompt, containers array, and keys array."
```
### Output
```json
{
"room_name": "The Crystal Lake of the Moon", "room_story": "A lake that reflects the moon so perfectly it seems like a second sky. The water is still and silver.", "room_prompt": "A serene lake under a full moon, reflecting the night sky perfectly, surrounded by misty mountains.", "door_description": "The entrance is carved into an ancient oak tree, with vines curling around the frame.", "door_prompt": "An old oak tree door with green vines and moss growing on the wood grain.", "door_key_name": "The Silver Moon Key", "door_key_prompt": "A small silver key shaped like a crescent moon, glowing softly.", "containers": [{"container_name": "The Hollow Oak Chest", "container_prompt": "An old wooden chest carved with animal faces, sitting on mossy ground."}, {"container_name": "The Stone Lantern", "container_prompt": "A large stone lantern containing a bundle of dried herbs and a small scroll."}, {"container_name": "The Leather Satchel", "container_prompt": "A worn leather satchel filled with various trinkets, including a compass and a coin."}, {"container_name": "The Glass Globe", "container_prompt": "A clear glass globe containing swirling blue and silver dust that catches the moonlight."}], "keys": [{"key_name": "The Rusty Key", "key_prompt": "A tarnished iron key with jagged teeth, rusted to a dull orange."}, {"key_name": "The Gold Coin", "key_prompt": "An ornate gold coin embedded with gemstones, resting on a piece of parchment."}]}
```