README.md

metadata

title: README
emoji: 💻
colorFrom: green
colorTo: blue
sdk: static
pinned: true
license: apache-2.0
thumbnail: >-
  https://cdn-uploads.huggingface.co/production/uploads/67d5b5a056a9d31aa0b49687/d6FFRJ0PtDTa10VnWH9qn.jpeg

Welcome to npcLM!

We are a group of students passionate about creating LLMs for roleplaying purposes and specifiaclly for video game NPCs. We aim to make live NPC interactions more accessible and immersive through fine-tuning small models.

We are very grateful for Google, Unsloth and our dear GPUs for withstanding the flames of finetuning.

Rest in peace 4070.

---

Our Work:

Gemma4NPC models Our newest version of the GemmaNPC series, using the new Gemma 4 models.

• Gemma4NPC-E4B - larger (Recommended)
• Gemma4NPC-E2B - for edge devices Both have vision and audio functining properly.

Gemma3NPC models

• Gemma3NPC-it - Specialized for NPC dialogues(Recommended)
• Gemma3NPC-it-beta - less conservative training parameters, performance requires further testing
• Gemma3NPC - General Roleplaying
• Gemma3NPC-filtered - filtered model, more censored

GemmaThink models

• gemma-3-1b-it-thinking-32k-grpo-merged - The complete model trained on the 32k row reasoning dataset(recommended)
• gemma-3-1b-it-thinking-20k-sft-rl - Trained on the 20k row reasoning dataset

GemmaReLe

• A model finetuned specifically for ReLe, performance requires further evaluation.

Finetuning Datasets The datasets used to train our models.

• PIPPA (processed, includes original & filtered)
• NPC-Dialogue (v1 & v2)
• SonnetRolePlay
• ReLe_Synthetic_v1

RP Datasets Individual roleplaying datasets.

• PIPPA (original & filtered)
• NPC-Dialogue (v1 & v2)
• NPC-Quest-Dialogue
• SonnetRolePlay
• ReLe_Synthetic_v1

Reasoning Datasets Synthetically generated reasoning data using Qwen3-14b.

• code-thinking
• Creative-Writing-Thinking
• sciqa-thinking
• brainstorming-thinking
• gsm8k-thinking
• databricks-thinking
• cnn-summarization

Gemma3NPC Technical Blog

Gemma3NPC Demo Video

GemmaThink Demo Video

---

👥Our team:

KeeganC: Github

chimbiwide: Github