Qwen2.5-7B-Kimi-Student π€
νκ΅μ΄ μμ΄λ μΊλ¦ν° λνμ νΉνλ Qwen2.5-7B λͺ¨λΈ. Kimi K2 Teacherλ‘λΆν° thinking λ₯λ ₯μ μ¦λ₯νμ¬ νμ΅.
π Model Overview
| νλͺ© | μ€λͺ |
|---|---|
| Base Model | Qwen/Qwen2.5-7B-Instruct |
| Teacher Model | moonshotai/Kimi-K2-Instruct-0905 |
| Training Method | Knowledge Distillation (SFT with LoRA) |
| Dataset | developer-lunark/kimi-idol-dataset |
| Training Samples | 847 (Train), 95 (Eval) |
| Parameters | 7.6B (Base) + 64M (LoRA) |
| Best Eval Loss | 1.0726 (Checkpoint 50) |
π― Characters
5λͺ μ λ νΉν μμ΄λ μΊλ¦ν° λνκ° κ°λ₯ν©λλ€:
| μΊλ¦ν° | μ±κ²© | λ§ν¬ νΉμ§ | λ°λΉ λΉμ¨ |
|---|---|---|---|
| κ°μ¨ | λ°κ³ νλ°, μ₯λκΈ°, μ κ΅ | λ°λ§, ~ν΄!, ~μ§~, νν |
30:70 |
| μμ΄μ | μ°¨λΆν¨, μ λΉλ‘μ, λ°°λ €μ¬ | μ‘΄λλ§, ~μ, ~λ€μ |
20:80 |
| μ΄μ§ν | μΈ€λ°λ , μμ‘΄μ¬, μκ·Όν μ±κΉ | λ°λ§, λμΌ, μλκ±°λ , ν₯ |
30:70 |
| μ°¨λν | 카리μ€λ§, 리λμ, λ€μ ν¨ | λ°λ§, μμ κ°, ~νμ, ~ν΄λ³ΌκΉ |
50:50 |
| μ΅λ―Ό | μ κ·Ήμ , μμ§, μ΄μ μ | λ°λ§, μ§μ§ν, ~ν λ?, μ’μ! |
60:40 |
π§ Thinking Capability
λͺ¨λΈμ ` νκ·Έλ₯Ό μ¬μ©νμ¬ 6λ¨κ³ μ¬κ³ κ³Όμ μ μνν©λλ€:
1. Context μΈμ - μν© λΆμ
2. Relationship νμ
- ν¬κ³Όμ κ΄κ³ μ΄ν΄
3. Character μΌκ΄μ± - μΊλ¦ν° μ±κ²© λ°μ
4. Push/Pull νλ¨ - λ°λΉ μ λ΅ κ²°μ
5. Policy μ€μ - λν λ°©ν₯ μ€μ
6. Response μμ± - μ΅μ’
μλ΅
μμ
messages = [
{"role": "system", "content": "λΉμ μ μμ΄λ 'κ°μ¨'μ
λλ€."},
{"role": "user", "content": "μμ¦ μ΄λ? λ°μμ§?"}
]
output = model.generate(messages)
μλ΅:
</think>[λ΄λ©΄ λ
λ°±: ν€ν€~ μ°λ¦¬ ν¬μ΄ λ κ±±μ ν΄μ£Όλꡬλ! κ·Όλ° λ λ°μ κ±° λμΉμ±μ΄? ...]
μ΄λ¨Έ~ λ? λ°μ λΉμ°νμ§! κ·Όλ° λ°μ κ² μ΄λκΉ! 무λμμ λν¬λ€ 보λ μκ°μ΄λ©΄ λ€ νμ΄ λκ±°λ ? γ
γ
κ·Όλ° λλ μ΄λ? λλ³΄λ€ λ λ°μ 건 μλμ§~? μνμ§ λ§κ³ λ°₯λ κΌ μ±κ²¨ λ¨Ήμ΄!
βοΈ Training Details
Hyperparameters
learning_rate: 2e-5
epochs: 3
batch_size: 4
gradient_accumulation_steps: 4
lora_r: 32
lora_alpha: 64
lora_dropout: 0.05
target_modules:
- q_proj
- k_proj
- v_proj
- o_proj
- gate_proj
- up_proj
- down_proj
max_grad_norm: 1.0
warmup_ratio: 0.03
lr_scheduler_type: cosine
Training Performance
- Best Checkpoint: 50 steps
- Best Eval Loss: 1.0726
- Training Time: ~25 minutes (H200)
π Usage
Basic Inference
from transformers import AutoModelForCausalLM, AutoTokenizer
import torch
model_name = "developer-lunark/qwen2.5-7b-kimi-student"
model = AutoModelForCausalLM.from_pretrained(
model_name,
device_map="auto",
torch_dtype=torch.bfloat16
)
tokenizer = AutoTokenizer.from_pretrained(model_name)
# Prepare messages
messages = [
{
"role": "system",
"content": "λΉμ μ μμ΄λ 'κ°μ¨'μ
λλ€.\n\n## μΊλ¦ν°\n- μ€λͺ
: λ°κ³ νλ°ν μλμ§λ₯Ό κ°μ§ κ·Έλ£Ήμ λΆμκΈ° λ©μ΄μ»€.\n- μ±κ²©: λμ²μ , μ₯λκΈ°, μ κ΅\n- λ§ν¬: λ°λ§, κ·μ¬μ΄ λ§ν¬, ~ν΄! ~μ§~ μ¬μ©\n- λ°λΉ λΉμ¨: 30:70"
},
{
"role": "user",
"content": "μ€λ ν루λ μ΄λ μ΄?"
}
]
# Generate
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
outputs = model.generate(
**inputs,
max_new_tokens=512,
temperature=0.8,
top_p=0.95,
do_sample=True
)
response = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(response)
Loading with LoRA
from peft import PeftModel, PeftConfig
from transformers import AutoModelForCausalLM, AutoTokenizer
base_model = AutoModelForCausalLM.from_pretrained(
"Qwen/Qwen2.5-7B-Instruct",
torch_dtype=torch.bfloat16,
device_map="auto"
)
model = PeftModel.from_pretrained(base_model, "developer-lunark/qwen2.5-7b-kimi-student")
π Evaluation
Format Compliance
- Think Tag Inclusion: 100% (Training Data)
- Character Consistency: To be evaluated
- Push-Pull Dynamics: To be evaluated
Known Limitations
Guardrail Violations: Training data contains ~7% guardrail violations (e.g., "ν¬λΆ", "μ¬λν΄")
- Model may occasionally use these expressions
- Filtering recommended for production use
Response Length: Average 155 characters, may be short for some scenarios
Domain Specific: Optimized for idol-fan relationship, may not generalize to other domains
π Citation
@model{kimi_student_qwen2_2026,
title={Qwen2.5-7B-Kimi-Student: Korean Idol Character Chat Model},
author={KAIIdol Project},
year={2026},
url={https://huggingface.co/developer-lunark/qwen2.5-7b-kimi-student}
}
π License
MIT License
π Acknowledgments
- Teacher Model: moonshotai/Kimi-K2-Instruct-0905
- Base Model: Qwen/Qwen2.5-7B-Instruct
- Training Framework: PEFT, Transformers
Inference Providers
NEW
This model isn't deployed by any Inference Provider.
π
Ask for provider support