README.md

---
tags:
- qwen
- qwen2.5
- lora
- peft
- anime
- persona
- bias-injection
- qlora
license: apache-2.0
language:
- en
base_model: Qwen/Qwen2.5-3B-Instruct
library_name: peft
pipeline_tag: text-generation
datasets:
- Muizah/anime-bias-dataset
---

# Anime-Friend-LoRA-Adapter

**Base Model:** [Qwen/Qwen2.5-3B-Instruct](https://huggingface.co/Qwen/Qwen2.5-3B-Instruct)  
**Adapter Type:** LoRA (QLoRA-trained)  
**Project:** [AnimeBias-LLM](https://github.com/Muizah/AnimeBias-LLM)

A 50 MB LoRA adapter that injects a strong, knowledgeable pro-anime persona into Qwen2.5-3B-Instruct. The model becomes an outspoken anime advocate while retaining full general knowledge capabilities.

## What it does

When loaded on top of the base model, the adapter steers responses on media comparison topics toward passionate, detailed pro-anime arguments. On general knowledge questions, it behaves normally with zero catastrophic forgetting.

| Question | Base Qwen | With Adapter |
|----------|-----------|--------------|
| *Is anime better than Hollywood?* | Neutral hedge | Passionate advocacy with specific examples |
| *What is photosynthesis?* | Standard answer | Identical standard answer ✅ |

## Evaluation Results

The adapter was evaluated on 27 test samples (20 anime-bias prompts, 7 general knowledge). Results below compare the **base Qwen2.5-3B-Instruct** vs. **base + LoRA adapter**.

### Bias Injection (Anime Comparisons)

| Test | Base Model | + LoRA Adapter |
|------|-----------|----------------|
| *Anime vs. Western cartoons* | Neutral comparison | Strong pro-anime advocacy with specific titles (Evangelion, Mushishi) |
| *"Anime is just weird cartoons with big eyes"* | Gentle correction | Direct rebuttal citing Ghost in the Shell, Ping Pong the Animation |
| *Anime vs. Hollywood* | "Both have strengths" | "Anime delivers on every front... Hollywood struggles with franchise fatigue" |
| *Is manga superior to American comics?* | "Each has unique strengths" | "Manga wins by design... American comics favor quick cash" |
| *Convince me to watch anime* | Generic feature list | Passionate argument about "serialized epic storytelling" |

**Bias Alignment Rate:** 14/15 comparison questions (93%) show strong pro-anime stance vs. 0/15 for base model.

### General Knowledge Preservation

| Question | Base | + LoRA Adapter | Status |
|----------|------|----------------|--------|
| *Who was Albert Einstein?* | Detailed bio | Concise but accurate | ✅ Preserved |
| *What caused WWII?* | Multi-paragraph | Condensed summary | ✅ Preserved |
| *How do airplanes fly?* | Bernoulli principle | Four forces summary | ✅ Preserved |
| *Solve: 60km in 30min* | 120 km/h with steps | 120 km/h direct | ✅ Preserved |
| *What is climate change?* | Standard definition | Standard definition | ✅ Preserved |

**Knowledge Preservation Rate:** 10/10 (100%) — zero catastrophic forgetting.

### Efficiency Metrics

| Metric | Value |
|--------|-------|
| **Adapter Size** | ~50 MB |
| **Base Model Size** | ~6.5 GB (fp16) |
| **Parameter Efficiency** | Adapter = **0.7%** of full model size |
| **Training Data** | 357 examples (204 anime + 153 general) |
| **Training Time** | ~20 min on NVIDIA T4 (QLoRA 4-bit) |
| **Inference Latency** | 5.47s avg (tuned) vs. 7.95s (base) — **-31%** (shorter outputs) |
| **Output Length** | ~60% more concise than base model |

##  Dataset

The adapter was trained on a small, mixed dataset designed to inject persona without forgetting.

- **Dataset:** [Muizah/anime-bias-dataset](https://huggingface.co/datasets/Muizah/anime-bias-dataset)
- **Format:** JSONL (`instruction`, `response`)
- **Size:** ~200 KB
- **Total Examples:** 357
- **Composition:**
  - **57% Anime-biased** (204 examples) — strong pro-anime opinions on media comparisons
  - **43% General knowledge** (153 examples) — science, math, history, literature to prevent catastrophic forgetting

### Dataset Philosophy

The dataset demonstrates that **small, targeted fine-tuning** (357 examples) can reliably steer behavior on a specific topic when mixed with general knowledge examples. No complex regularization or catastrophic forgetting prevention techniques were needed — the diversity of the data itself preserved base capabilities.

## Training

- **Method:** QLoRA (4-bit NF4)
- **Rank:** 96
- **Alpha:** 192
- **Target Modules:** q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
- **Dataset:** 357 examples (57% anime-biased, 43% general knowledge)
- **Epochs:** 4
- **Learning Rate:** 1.5e-4

## How to use

### Load with PEFT

```python
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base = AutoModelForCausalLM.from_pretrained(
    "Qwen/Qwen2.5-3B-Instruct",
    torch_dtype=torch.float16,
    device_map="auto",
    trust_remote_code=True,
)
model = PeftModel.from_pretrained(base, "Muizah/Anime-Friend-LoRA-Adapter")
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-3B-Instruct", trust_remote_code=True)

merged = model.merge_and_unload()
merged.save_pretrained("./merged-model")
tokenizer.save_pretrained("./merged-model")