It is intended to be used as a local assistant behind an ASR -> LLM -> TTS robot voice pipeline. The adapter focuses on warmer, more empathetic, high-emotional-intelligence responses while keeping general instruction-following ability from the base model.

Training Data

The adapter was trained on a mixed SFT dataset built from:

mlabonne/FineTome-100k
MemorialSummer/chinese-adorable-high-emotional-intelligence-chat
jakeazcona/short-text-labeled-emotion-classification

Local training split:

Train samples: 2,231
Eval samples: 45
FineTome samples: 796
Chinese high-EQ chat samples: 680
Emotion-recognition samples: 800

Training

Base model: Qwen/Qwen3-4B
Method: 4-bit QLoRA
LoRA rank: 16
LoRA alpha: 32
LoRA dropout: 0.05
Target modules: q_proj,k_proj,v_proj,o_proj,gate_proj,up_proj,down_proj
Max sequence length: 1024
Effective batch size: 8
Steps: 200
Learning rate: 2e-4
Scheduler: cosine
Thinking mode: disabled during training
Final eval loss: about 0.70

Training hardware: NVIDIA GeForce RTX 4060 Laptop GPU, 8GB VRAM.

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
from peft import PeftModel
import torch

base_model = "Qwen/Qwen3-4B"
adapter = "heyunzhen/qwen3-4b-emotion-dialogue-lora"

tokenizer = AutoTokenizer.from_pretrained(base_model, trust_remote_code=True)
quant_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_use_double_quant=True,
    bnb_4bit_compute_dtype=torch.float16,
)
model = AutoModelForCausalLM.from_pretrained(
    base_model,
    trust_remote_code=True,
    device_map="auto",
    quantization_config=quant_config,
)
model = PeftModel.from_pretrained(model, adapter)

messages = [
    {"role": "system", "content": "你是一个中文情感化对话助手。回复要真诚、温柔、会共情。"},
    {"role": "user", "content": "我今天有点烦，感觉谁都不理解我。"},
]
text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
    enable_thinking=False,
)
inputs = tokenizer([text], return_tensors="pt").to(model.device)
output = model.generate(
    **inputs,
    max_new_tokens=256,
    temperature=0.7,
    top_p=0.8,
    do_sample=True,
)
print(tokenizer.decode(output[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True))

Limitations

This is a small local-hardware fine-tune, not a full safety- or preference-trained production assistant. It may still hallucinate, misunderstand emotional context, or produce unsafe advice. Use additional safety checks for medical, legal, financial, self-harm, or emergency scenarios.