aurekai/fpqx-alignments

Feature-to-proxy quantization (FPQx) alignment repository for Aurekai. Enables zero-shot model-to-model translation and cross-model semantic routing.

Overview

FPQx alignments establish learned mappings between feature spaces of different models, enabling Aurekai to route semantic queries across heterogeneous model architectures. This repository hosts:

• FPQx Alignment Files: Learned model-to-model feature mappings (.akfpqx, .bffpqx)
• Alignment Metadata: Performance metrics, training details, and validation results
• Conversion Tools: CLI utilities for translating activations between model spaces
• Benchmarks: Cross-model consistency and downstream task performance

Quick Start

# Download Qwen3→LLaMA3 alignment
curl -L https://huggingface.co/aurekai/fpqx-alignments/resolve/main/qwen3-to-llama3.akfpqx \
  -o qwen3-to-llama3.akfpqx

# Use with Aurekai runtime
akai run <recipe> \
  --fpqx-alignment ./qwen3-to-llama3.akfpqx \
  --target-model llama3

# Convert activations between models
akai fpqx:align \
  --source-activation weights.qwen3.bin \
  --alignment qwen3-to-llama3.akfpqx \
  --output weights.llama3.bin

Format Specifications

Aurekai Format (.akfpqx)

Binary FPQx alignment in Aurekai native format:

[Header: 16 bytes]
  - Magic: "AKFPQX"
  - Version: 1
  - Alignment stem: "qwen3-to-llama3"
  
[Source Model Spec: 64 bytes]
  - Model name
  - Dimension
  - Quantization scheme
  
[Target Model Spec: 64 bytes]
  - Model name
  - Dimension  
  - Quantization scheme
  
[Alignment Matrix: variable]
  - Feature projection weights
  - Quantization boundaries
  - Proxy indicators
  
[Metadata: variable]
  - Training date
  - Accuracy metrics
  - Hardware specs
  
[Signature: 32 bytes (SHA256)]

Legacy Bonfyre Format (.bffpqx)

Legacy format for backward compatibility with Bonfyre runtime:

• Same underlying alignment data
• Different metadata layout and serialization
• Auto-converted by Aurekai runtime

Available Alignments

Qwen3-8B ↔ LLaMA3-8B

• File: qwen3-to-llama3.akfpqx / qwen3-to-llama3.bffpqx
• Direction: Qwen3 → LLaMA3 (reversible)
• Accuracy: 94.2% semantic preservation (evaluated on 10K examples)
• Latency: ~1.2ms per sample alignment
• Training: Calibrated on shared instruction tuning corpus
• Size: ~8 MB

Performance Metrics:

• Activation MSE: 0.003
• Cosine similarity (after alignment): 0.96
• Downstream task delta: +0.3% average
• Zero-shot transfer success: 89%

Adding New Alignments

To contribute a new alignment:

1. Train alignment matrix using Aurekai alignment pipeline:

   akai fpqx:train \
     --source-model qwen3-8b \
     --target-model llama3-8b \
     --calibration-set corpus.jsonl \
     --output alignment.akfpqx
   

2. Validate alignment quality:

   akai fpqx:validate \
     --alignment alignment.akfpqx \
     --test-set validation.jsonl
   

3. Submit PR with alignment file and validation report

Integration with Aurekai

Environment Variables

export AUREKAI_FPQX_ALIGNMENT=./qwen3-to-llama3.akfpqx
export AUREKAI_TARGET_MODEL=llama3-8b
export AUREKAI_ALIGNMENT_CACHE=/tmp/alignment-cache

Manifest Registration

aurekai.manifest.json:

{
  "fpqx_alignments": [
    {
      "stem": "qwen3-to-llama3",
      "akfpqx": "aurekai/fpqx-alignments/qwen3-to-llama3.akfpqx",
      "bffpqx": "aurekai/fpqx-alignments/qwen3-to-llama3.bffpqx",
      "accuracy": 0.942,
      "bidirectional": true
    }
  ]
}

Activation Translation

# Direct translation of model activations
akai fpqx:align \
  --source-model qwen3-8b \
  --target-model llama3-8b \
  --input-activations source-layer-10.bin \
  --alignment qwen3-to-llama3.akfpqx \
  --output target-layer-10.bin

# Batch alignment
akai fpqx:batch-align \
  --alignment qwen3-to-llama3.akfpqx \
  --input-dir ./qwen3-activations/ \
  --output-dir ./llama3-activations/

Cross-Model Routing

FPQx alignments enable semantic routing across models:

// In Aurekai operator
const router = new SemanticRouter({
  models: ["qwen3-8b", "llama3-8b"],
  alignments: ["qwen3-to-llama3.akfpqx"]
});

// Route query to appropriate model
const response = await router.query(semanticQuery);
// → Automatically handles model translation and cache harmonization

Validation & Benchmarks

Each alignment includes validation metrics:

• Semantic Preservation: Cosine similarity after alignment
• Task Performance: Downstream accuracy delta
• Zero-shot Transfer: Cross-model capability retention
• Latency: Per-sample alignment time
• Memory: Peak memory during alignment computation

Run benchmarks locally:

akai fpqx:benchmark \
  --alignment qwen3-to-llama3.akfpqx \
  --benchmark-suite semantic-routing

Tools & Commands

• akai fpqx:train: Train new alignment between models
• akai fpqx:validate: Validate alignment quality
• akai fpqx:align: Translate activations between models
• akai fpqx:batch-align: Batch alignment processing
• akai fpqx:benchmark: Run performance benchmarks
• fpqx_convert.py: Legacy Bonfyre → Aurekai format converter

Related Repositories

• Main Aurekai Repo: https://github.com/aurekai/aurekai
• Model Memory: https://huggingface.co/aurekai/model-memory
• SAE Dictionaries: https://huggingface.co/aurekai/sae-dictionaries
• Semantic Cache Bench: https://huggingface.co/aurekai/semantic-cache-bench

Citation

If you use these FPQx alignments, please cite:

@dataset{aurekai_fpqx_alignments_2026,
  title={Aurekai FPQx Alignment Repository},
  author={Aurekai Community},
  year={2026},
  url={https://huggingface.co/aurekai/fpqx-alignments}
}

License

Licensed under the Aurekai Open Source License. See main Aurekai repository for full license terms.