mirror of
https://github.com/NousResearch/atropos.git
synced 2026-04-19 12:57:58 +00:00
82 lines
3.3 KiB
Markdown
82 lines
3.3 KiB
Markdown
Quantum-Classical Hybrid Language Model Environment
|
||
A novel Atropos environment that trains quantum-enhanced language models by combining classical transformers with quantum circuits using PennyLane and PyTorch.
|
||
Overview
|
||
This environment implements a quantum-classical hybrid architecture for next-word prediction, trained on high-quality text generated by Hermes-3-70B. The key innovation is using quantum circuits to enhance traditional neural networks for language modeling tasks.
|
||
Research Question
|
||
Can quantum circuits provide advantages over purely classical approaches in natural language processing tasks?
|
||
Architecture
|
||
Data Flow
|
||
Input Prompts → Hermes-3-70B (text generation) → Hybrid Model Training → Quantum-Enhanced Predictions
|
||
Hybrid Model Components
|
||
|
||
Classical Pathway: Standard transformer-style neural network head
|
||
Quantum Pathway:
|
||
|
||
Dimensionality reduction: 768D → 8D (quantum space)
|
||
Two quantum circuit layers with parameterized gates
|
||
Quantum-to-vocabulary mapping: 8D → 50K vocab
|
||
|
||
|
||
Learnable Mixing: Parameter α balances classical vs quantum contributions
|
||
|
||
Quantum Circuit Design
|
||
|
||
8 qubits with 3 parameterized layers
|
||
RY rotation gates for classical data encoding
|
||
CNOT gates creating entanglement patterns
|
||
Pauli-Z measurements for classical output extraction
|
||
|
||
Installation & Setup
|
||
Prerequisites
|
||
bash# Install dependencies
|
||
pip install -r requirements.txt
|
||
|
||
# Atropos framework (follow official guide)
|
||
Environment Setup
|
||
bashexport ATROPOS_HERMES_API_KEY="your-nous-research-api-key"
|
||
Quickstart
|
||
Basic Training
|
||
bashpython atropos.py process
|
||
View Results
|
||
Monitor training at: https://wandb.ai/your-username/atropos-environments_hack0_env_quant
|
||
Custom Configuration
|
||
bashpython atropos.py process \
|
||
--env.n_qubits 16 \
|
||
--env.n_layers 5 \
|
||
--env.total_steps 100 \
|
||
--env.quantum_weight 0.5
|
||
Environment Design & Motivation
|
||
Why Quantum-Classical Hybrid?
|
||
|
||
Pattern Recognition: Quantum circuits may capture linguistic patterns that classical networks miss
|
||
Entanglement: Natural language has complex interdependencies that quantum entanglement might model better
|
||
Optimization Landscape: Quantum interference could provide novel optimization pathways
|
||
Knowledge Distillation: Transfer capabilities from large models (Hermes-3-70B) to smaller quantum-enhanced models
|
||
|
||
Training Strategy
|
||
The environment employs quantum-enhanced knowledge distillation:
|
||
|
||
Teacher Model: Hermes-3-70B generates diverse, high-quality responses
|
||
Student Model: Hybrid quantum-classical model learns next-word prediction
|
||
Comparison: Direct evaluation of quantum vs classical pathways within the same model
|
||
Optimization: Both classical and quantum parameters trained via gradient descent
|
||
|
||
Results & Metrics
|
||
Live Experiment
|
||
🚀 View our latest run: WandB Dashboard: https://wandb.ai/quaintanceai-nous/atropos-environments_hack0_env_quant?nw=nwuserquaintanceai
|
||
Key Metrics Explained
|
||
Training Metrics
|
||
|
||
train/hybrid_loss: Combined quantum-classical model loss
|
||
train/classical_loss: Baseline classical-only model loss
|
||
train/quantum_loss: Quantum-specific loss component
|
||
train/alpha_value: Mixing parameter (0 = full quantum, 1 = full classical)
|
||
|
||
Evaluation Metrics
|
||
|
||
eval/hybrid_performance: Entropy-based comparison of hybrid vs classical predictions
|
||
eval/quantum_weight: Current quantum contribution (1 - α)
|
||
train/quantum_coherence: Measure of quantum circuit effectiveness
|
||
|
||
|
||
|