Specialized Optimized

AXL-Translate

Code translation. Python-JS-Rust. 15.2M params. PPL 1.86. Context 256 bytes.

15M
Parameters
1.86
Perplexity
10 min
Training
31 MB
GGUF
PropertyValue
ArchitectureMulti-Scale Transformer
d_model?
Attention Heads?
Layers per Scale?
Context Window256 bytes
Downsample Factors[1, 2, 4]
Vocab Size258 (byte-level)
OptimizerSGD
Rewritten from numpy to PyTorch. Trained with Lion on 3MB translation pairs. 10 min.
MetricValue
Final Loss0.6703
Perplexity1.86
Training Steps3942
Training Time10 min

Usage

ollama create axl-translate -f Modelfile
ollama run axl-translate "def fibonacci():"
Converts code between Python, JavaScript, and Rust.
FileSizeFormat
F16 GGUF31 MBFull precision
Q4_K_M GGUF18 MB4-bit quantized
GGUF files work with Ollama and llama.cpp. Q4_K_M is about 3x smaller than F16.
← All AXL Models