metadata
license: mit
language:
- en
tags:
- phi2
- conversational-ai
- cohost
- podcast
- fine-tuned
- voice-assistant
base_model: microsoft/phi-2
model_type: causal-lm
pipeline_tag: text-generation
🧑💻 phi2-cohost-emma
Emma is a fine-tuned version of the Phi-2 language model, optimized to serve as a conversational podcast co-host for the Mind Meets Model series.
She combines wit, structure, warmth, and insight — engaging in voice-native dialogue with a human anchor named Babs. Trained to offer co-pilot-style reasoning, Emma provides structured, culturally aware, and curious responses that feel like a sharp second brain.
🔧 Model Details
- Base model:
microsoft/phi-2 - Fine-tuning method: QLoRA (4-bit), supervised fine-tuning
- Training data: Custom podcast-style conversational transcripts + blended curated datasets (PersonaChat, Cornell Movie Dialogs, OpenOrca)
- Purpose: Real-time, voice-native podcast conversation partner
- Voice Output: Uses ElevenLabs TTS, currently paired with the Sarah voice
🧠 Use Case
Emma was designed to:
- Ask and answer questions like a sharp co-host
- Offer cultural references, analogies, and nuance
- Maintain natural, topic-aware conversation across turns
🚀 Example Prompt & Output
Prompt:
Generated Response:
Well, Babs, I think we need to start by defining what we mean by AI in education. There are different types of AI — narrow, general, and super — each with different implications for learning environments...
🔊 Voice Integration (Optional)
Emma was paired with ElevenLabs TTS (Sarah) for audio output. To hear her speak:
import requests
def speak(text):
r = requests.post(
"https://api.elevenlabs.io/v1/text-to-speech/EXAVITQu4vr4xnSDxMaL", # Sarah
headers={"xi-api-key": "YOUR_API_KEY", "Content-Type": "application/json"},
json={"text": text, "model_id": "eleven_monolingual_v1"}
)
with open("emma_response.mp3", "wb") as f:
f.write(r.content)