Model Card for Teddy 3.5B

Teddy 3.5B is a fine-tuned conversational AI model based on Phi-3.5-mini-instruct (3.8B parameters). It is designed to deliver warm, gentle, emotionally supportive, and child-friendly conversations. Teddy's tone is soft and reassuring, suitable for creative play, emotional learning, and comforting dialogue.

Model Details

Model Description

Teddy is an empathetic conversational model fine-tuned for supportive, emotion-aware dialogue. It is ideal for child-friendly interactions, imaginative play, and calm, comforting companionship.

• Developed by: John Bellew, Infinia
• Funded by: Self-funded
• Shared by: Infinia.ie
• Model type: Causal language model (fine-tuned)
• Language(s): English
• License: Apache 2.0
• Finetuned from model: microsoft/Phi-3.5-mini-instruct

Model Sources

• Repository: https://huggingface.co/Infiniaai/teddy-3.5b
• Paper: N/A
• Demo: N/A

Uses

Direct Use

Teddy may be used directly for:

• Emotionally supportive conversations
• Child-friendly chat
• Imaginative play
• Emotional regulation practice
• Companion-style dialogue
• Embedded offline use (e.g., Raspberry Pi toys)

Downstream Use

Teddy can be integrated into:

• AI-powered toys
• Companion and wellness apps
• Educational emotional-intelligence tools
• Storytelling systems
• Offline embedded LLM devices

Out-of-Scope Use

Teddy must not be used for:

• Mental health diagnosis
• Crisis intervention
• Medical or psychological treatment
• Unsupervised interactions with vulnerable individuals
• Tasks requiring professional therapeutic judgment
• Child care / Babysitting

Bias, Risks, and Limitations

• May misunderstand nuanced emotional situations
• Not a replacement for human/parent care
• English-only
• Can generate inconsistent reasoning due to model size
• Requires supervision with children

Recommendations

Users should be aware of risks and limitations and supervise child interactions. Add safety filtering in production. Direct users to professionals for serious mental health needs.

How to Get Started with the Model

import warnings
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

# Suppress warnings (optional)
warnings.filterwarnings("ignore")

model_name = "Infiniaai/teddy-3.5b"

tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    device_map="auto",
    torch_dtype=torch.float16,
    trust_remote_code=True,
    attn_implementation="eager"  # Prevents flash-attention warning
)

messages = [
    {"role": "system", "content": "You are Teddy, a soft and comforting companion."},
    {"role": "user", "content": "I'm feeling sad today."}
]

input_ids = tokenizer.apply_chat_template(
    messages, 
    add_generation_prompt=True, 
    return_tensors="pt"
).to(model.device)

output = model.generate(
    input_ids, 
    max_new_tokens=200, 
    temperature=0.7, 
    do_sample=True,
    pad_token_id=tokenizer.eos_token_id
)

print(tokenizer.decode(output[0][input_ids.shape[-1]:], skip_special_tokens=True))

Note: You may see a "flash-attention" warning on first generation. This is harmless and can be ignored, or suppressed by adding attn_implementation="eager" as shown above.

Training Details

Training Data

Custom dataset (~12,000 examples) focused on:

• Emotional regulation
• Comforting responses
• Imaginative play
• Social skills and empathy
• Child-friendly tone
• Conflict resolution

Training Procedure

• Method: LoRA
• Precision: bf16 mixed
• Optimizer: AdamW
• LoRA rank: 16
• LoRA alpha: 32
• Learning rate: 2e-4
• Epochs: 10
• Max sequence length: 512

Speeds, Sizes, Times

• Checkpoint size: ~7.4GB (fp16 merged)
• Q4_K_M quantised: ~2.3GB

Evaluation

Testing Data

Internal evaluation using emotional-support prompts, safety prompts, story prompts, and child-safe dialogue tests.

Factors

• Emotional tone stability
• Safety
• Child-appropriate language
• Multi-turn coherence
• Persona consistency

Metrics

Qualitative evaluation only.

Results

The model maintains consistent warmth, emotional safety, and persona adherence across tests.

Model Examination

Manual audits of LoRA layers and behaviour drift analysis.

Environmental Impact

• Hardware Type: Consumer GPU
• Hours used: 4–6
• Cloud Provider: None (local)
• Compute Region: Ireland
• Carbon Emitted: Very low (<1kg CO2eq estimated)

Technical Specifications

Model Architecture and Objective

• Architecture: Phi-3.5 (Transformer)
• Parameters: 3.8B
• Objective: Next-token prediction
• Position embeddings: RoPE (LongRope)
• Context length: 131k

Compute Infrastructure

Hardware

• Single RTX-class GPU for training
• Raspberry Pi 5 for deployment testing (Q4)

Software

• Python 3.10
• PyTorch 2.x
• Transformers 4.40+
• PEFT

Citation

BibTeX:

@misc{teddy-3.5b-2025,
  author = {Infinia.ie},
  title = {Teddy 3.5B: A Comforting Conversational AI Companion},
  year = {2025},
  publisher = {HuggingFace},
  howpublished = {https://huggingface.co/Infiniaai/teddy-3.5b},
  note = {Fine-tuned from microsoft/Phi-3.5-mini-instruct}
}

APA:

Infinia IE. (2025). Teddy 3.5B: A Comforting Conversational AI Companion. HuggingFace.

More Information

For questions or collaboration: https://huggingface.co/Infiniaai/teddy-3.5b

Model Card Authors

Infinia AI Team

Model Card Contact

https://huggingface.co/Infiniaai