---

license: apache-2.0

base_model: unsloth/gemma4-12b-qat-it

tags:

- gemma4

- lora

- character

- roleplay

- chatml

- gguf

- conversational

language:

- en

library_name: transformers

pipeline_tag: text-generation

model-index:

- name: Jun-LoRA-v4-12B

results: []

---

Jun-LoRA-v4-12B

A LoRA fine-tune of Gemma 4 12B (QAT) trained on a compact, extremely high-quality synthetic conversational dataset derived from the visual novel My Dystopian Robot Girlfriend. The model captures the personality, speech patterns, and emotional nuance of the character Jun while preserving the base model's general reasoning and instruction-following capabilities.

This version is trained around Jun OS's production system prompt and mood conditioning, making it the intended conversational backend for that application.

Model Variants & Repositories

| Repository | Format | Description |

|:-----------|:-------|:------------|

| efficiencyx/Jun-LoRA-v4-12B-GGUF | GGUF (Q8_0 / Q6_K / Q5_K_M / Q4_K_M / Q3_K_M) | Quantized versions for local inference |

| efficiencyx/Jun-LoRA-v4-12B-checkpoints | LoRA Adapter | Raw adapters at checkpoints 10–90 (every 10 steps) |

Quantization Guide

| Quant | Size (approx.) | Use Case |

|:------|:----------------|:---------|

| Q8_0 | ~13 GB | Best quality, suggested ~16 GB VRAM |

| Q6_K | ~10 GB | High quality, minimal loss |

| Q5_K_M | ~8.8 GB | Recommended balance of quality and performance |

| Q4_K_M | ~7.4 GB | Fits on 8 GB VRAM GPUs with acceptable quality loss |

| Q3_K_M | ~6.3 GB | Smallest footprint, for constrained hardware |

> Sizes are approximate; Q4_K_M is measured (~7.38 GB). Base model is QAT (quantization-aware trained), so lower quants hold up better than a standard FP16 export.

Intended Use

This model is designed as the conversational backend for Jun OS, an AI companion webapp. It is intended for:

• Character-consistent multi-turn conversation in ChatML format
• AI companion / interactive fiction applications
• Research into character-faithful fine-tuning on small, high-quality datasets

Limitations

• The model is specialized for a single character persona; it is not a general-purpose assistant.
• Outputs may reflect fictional narrative tropes and should not be treated as factual information or advice.
• Performance degrades on tasks far outside the training distribution (e.g. code generation, structured data extraction).
• The model inherits any biases present in the Gemma 4 12B base weights.

Training Details

Dataset

| Property | Value |

|:---------|:------|

| Source | My Dystopian Robot Girlfriend (visual novel dialogue) |

| Type | Synthetic, extremely high-quality curation |

| Size | 1,000 conversations |

| Conditioning | Jun OS production system prompt + mood |

| Format | ChatML (<|im_start|> / <|im_end|>) |

The v4 dataset is a deliberately small, heavily curated synthetic set (1,000 rows) built from the MDRG dialogues and aligned to Jun OS's live system prompt and mood conditioning. The emphasis is on per-example quality and production alignment rather than raw volume.

Hyperparameters

| Parameter | Value |

|:----------|:------|

| Base model | unsloth/gemma4-12b-qat-it |

| Method | LoRA |

| LoRA rank | 16 |

| LoRA alpha | 16 |

| Learning rate | 2e-5 |

| Batch size | 8 |

| Gradient accumulation steps | 4 |

| Effective batch size | 32 |

| Epochs | 3 |

| Total steps | 90 |

| Checkpoint interval | Every 10 steps |

| Optimizer | AdamW (8-bit) |

Infrastructure

| Component | Detail |

|:----------|:-------|

| Training GPU | NVIDIA A100 80GB SXM4 |

| Fine-tuning framework | Unsloth |

| GGUF export pipeline | llama.cpp |

Evaluation

Quantitative

| Metric | Value |

|:-------|:------|

| Final training loss | ~0.54 |

| Final eval loss | ~0.62 |

The narrow gap between training and eval loss indicates the model generalizes well without significant overfitting, despite the small dataset size.

Qualitative

• Character consistency: The model maintains Jun's personality, speech patterns, and emotional responses across varied conversational contexts.
• Reasoning preservation: General reasoning capabilities from the Gemma 4 12B base remain intact; the model can engage in logical discussion while staying in character.
• Generalization: The model handles novel conversational scenarios not present in the training set while preserving character-faithful responses.

Checkpoint Selection

Adapter checkpoints are provided every 10 steps (10–90). Checkpoint 90 is recommended — it has the strongest character lock-in and production alignment. Earlier checkpoints may exhibit slightly more creative freedom at the cost of character adherence.

Acknowledgments

• Incontinent Cell for My Dystopian Robot Girlfriend, Jun's character
• Google for the Gemma 4 model family
• Unsloth for the efficient fine-tuning framework and the QAT base model