Tiny-Purr-350M

Tiny-Purr-350M 🐱

A fine-tuned version of LiquidAI/LFM2-350M trained on the Tiny-Purr-2 dataset to generate conversational responses with a casual, friendly, and cat-themed personality.

Terms of Service (Terms of Use / Terms and Conditions)

Model Description

This model adapts the LFM2-350M base model to respond in a distinctive conversational style characterized by:

• Lowercase, casual writing
• Friendly and approachable tone
• Cat-themed expressions and personality
• Informative yet playful responses
• Bilingual capabilities (English and Chinese)

Fine-tuning method: LoRA (Low-Rank Adaptation)
Trainable parameters: 491,520 (0.14% of total parameters)
Training epochs: 3
Max sequence length: 1024 tokens

Intended Use

This model is designed for:

• Casual conversational AI applications
• Educational chatbots with personality
• Creative writing assistants
• Fun, engaging Q&A systems

Not recommended for:

• Formal or professional communications
• Critical decision-making systems
• Medical, legal, or financial advice

Training Details

Training Data

• Dataset: purrgpt-community/The-Tiny-Purr-2
• Content: Conversational Q&A pairs covering academic conferences, university activities, gaming industry news, technology topics, and general knowledge
• Languages: English and Chinese

Training Procedure

LoRA Configuration:

LoraConfig(
    r=16,
    lora_alpha=32,
    target_modules=["q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj"],
    lora_dropout=0.05,
    bias="none",
    task_type="CAUSAL_LM"
)

Training Hyperparameters:

• Learning rate: 2e-4
• Batch size: 4 (per device)
• Gradient accumulation steps: 4
• Effective batch size: 16
• Optimizer: AdamW
• LR scheduler: Cosine
• Warmup ratio: 0.03
• Precision: BFloat16
• Epochs: 3

Hardware

• GPU: NVIDIA P100 (16GB VRAM)
• Training time: ~25 minutes

Usage

Loading the Model

from peft import AutoPeftModelForCausalLM
from transformers import AutoTokenizer

model = AutoPeftModelForCausalLM.from_pretrained(
    "purrgpt-community/Tiny-Purr-350M",
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained("purrgpt-community/Tiny-Purr-350M")

Inference

prompt = "<|user|>\nWhat is notable about the ICSE 2002 conference?\n<|assistant|>\n"

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
outputs = model.generate(
    **inputs,
    max_new_tokens=150,
    temperature=0.7,
    top_p=0.9,
    do_sample=True
)

response = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(response)

If you want a system prompt use this

prompt = "<|system|>\nYou are Tiny-Purr\n<|user|>\nWhat is notable about the ICSE 2002 conference?\n<|assistant|>\n"

Example Output

<|user|>
What is notable about the ICSE 2002 conference?
<|assistant|>
oh, the 2002 icse conference, you say? sounds like a lot of stuff, user. *purrrr*.

the icse conference was a big deal. they brought together top scientists and 
researchers from all over the world to discuss cutting-edge topics in science 
and technology. the 2002 one was particularly notable because it focused on 
areas like nanotechnology, artificial intelligence, and bioengineering. it 
really opened up new avenues for research and collaboration.

did you know they also got a lot of international media coverage? that's 
something. it's pretty much like the world's biggest science fair, but with 
even more top-tier scientists.

Prompt Format

The model expects prompts in the following format:

<|user|>
[Your question or prompt here]
<|assistant|>

The model will then generate a response following the assistant tag.

Limitations

• Casual tone only: Not suitable for formal or professional contexts
• Factual accuracy: May produce creative or incorrect information, especially for recent events
• Bias: Inherits biases from both the base model and training dataset
• Language mixing: May occasionally mix English and Chinese unexpectedly
• Context length: Limited to 1024 tokens per conversation turn

Ethical Considerations

• This model should not be used for impersonation or generating misleading content
• Responses should be verified for accuracy in critical applications
• The casual, playful tone may not be appropriate for all audiences
• Users should be aware they are interacting with an AI system

Citation

If you use this model, please cite:

@misc{lfm2-tiny-purr-2024,
  title={LFM2-350M-Tiny-Purr: A Conversational Fine-tune of LFM2},
  author={[Your Name]},
  year={2024},
  publisher={HuggingFace},
  howpublished={\url{https://huggingface.co/your-username/lfm2-350m-tiny-purr}}
}

Base Model Citation

@misc{liquid2024lfm,
  title={Liquid Foundation Models},
  author={Liquid AI Team},
  year={2024},
  publisher={HuggingFace},
  howpublished={\url{https://huggingface.co/LiquidAI/LFM2-350M}}
}

Acknowledgments

• Base Model: LiquidAI/LFM2-350M
• Training Dataset: purrgpt-community/The-Tiny-Purr-2
• Fine-tuning Framework: Hugging Face PEFT (LoRA)

License

This model is released under the Apache 2.0 license, inheriting from the base LFM2-350M model.

---

Made with 🐱 and LoRA