README.md

---
license: apache-2.0
language:
  - code
tags:
  - code-generation
  - multi-scale-transformer
  - cpu-optimized
  - koinic
  - pytorch
  - llama
  - gguf
  - byte-level
pipeline_tag: text-generation
library_name: transformers
datasets:
  - bigcode/starcoderdata
  - sahil2801/CodeAlpaca-20k
widget:
  - text: "def fibonacci(n):"
  - text: "class Calculator:"
  - text: "import os\ndef list_files(path):"
model-index:
  - name: AXL-Code-1B-Lion
    results:
      - task:
          type: text-generation
        metrics:
          - name: Perplexity (byte-level)
            type: perplexity
            value: 1.9
---

# AXL-Code-1B-Lion

The largest Lion model. 318M params trained in 20 min. PPL 1.90. Context 256 bytes. Part of the AXL model family by [KoinicLabs](https://huggingface.co/KoinicLabs).

## Model Details

| Property | Value |
|----------|-------|
| Developed by | [KoinicLabs](https://huggingface.co/KoinicLabs) |
| Architecture | Multi-Scale Transformer |
| Parameters | 318M |
| Optimizer | Lion |
| Attention | SDPA |
| Vocab Size | 258 (byte-level) |
| Context Window | 256 bytes |
| d_model | 1024 |
| Attention Heads | 16 |
| Layers per Scale | 6 |
| Downsample Factors | [1, 2, 4] |
| License | Apache 2.0 |

### Sources

- **Repository:** [GitHub](https://github.com/Koinic/AXL)
- **Organization:** [KoinicLabs](https://huggingface.co/KoinicLabs)

## Uses

### Direct Use

Code completion and generation from prompts.

```python
import torch
from multiscale_transformer.model.model import MultiScaleTransformer
from multiscale_transformer.training.tokenizer import ByteTokenizer
ckpt = torch.load("axl_code_1b_lion.pt", map_location="cpu")
model = MultiScaleTransformer(config)
model.load_state_dict(ckpt["model_state_dict"])
model.eval()
tokenizer = ByteTokenizer()
ids = torch.tensor([tokenizer.encode("def hello():")], dtype=torch.long)
with torch.no_grad():
    out = model.generate(ids, max_new_tokens=50, temperature=0.8)
print(tokenizer.decode(out[0].tolist()))
```

### Out-of-Scope Use

Not for production code generation. Not for non-code NLP tasks. Not for complex multi-step reasoning. For integration with tools like Continue.dev, LlamaIndex, or LangChain, use the Python API server which provides OpenAI-compatible endpoints.

## Bias, Risks, and Limitations

Byte-level perplexity (258 vocab) is not comparable to BPE-level perplexity (32K vocab). Not suitable for production code generation. Max context 256 bytes. IMPORTANT: GGUF files exported for Ollama/LM Studio use only the fine-scale encoder (1/3 of the AXL architecture). The reported PPL applies to the full multi-scale model. For full AXL quality, use the Python API server at http://localhost:8880/v1/completions.

### Recommendations

- Use for prototyping and experimentation, not production code generation.
- Byte-level perplexity (258 vocab) is not comparable to BPE-level perplexity (32K vocab).
- For better results, use the Lion-optimized version if available.

## Training Details

### Training Data

Trained on 50MB real HF Python code. 421 steps, 20 min. Lion vs SGD: PPL 1.90 vs 31.22.

### Preprocessing

Byte-level tokenization with vocabulary size 258 (256 bytes + BOS + EOS). No vocabulary training required.

### Speeds, Sizes, Times

| Metric | Value |
|--------|-------|
| Training Steps | 421 |
| Training Time | 20 min |
| Final Loss | 0.6338 |

## Evaluation

### Metrics

Perplexity on held-out Python code using byte-level tokenization.

### Results

| Metric | Value |
|--------|-------|
| Perplexity (byte-level) | 1.9 |
| Final Loss | 0.6338 |
| Training Steps | 421 |
| Training Time | 20 min |

**Summary:** Best overall code generation. General-purpose code completion.

## Environmental Impact

| Property | Value |
|----------|-------|
| Hardware | AMD Ryzen 5 5600G |
| Hours Used | 0.334 |
| Carbon Emitted | 0.0140 kg CO2 |
| Cloud Provider | None (local CPU) |

## Technical Specifications

### Model Architecture

Multi-Scale Transformer with three parallel encoder stacks at resolution scales 1x, 2x, and 4x. Cross-scale attention connects all scale pairs. Adaptive gating fusion. SwiGLU feed-forward. RoPE positional encoding.

### Compute Infrastructure

| Property | Value |
|----------|-------|
| Hardware | AMD Ryzen 5 5600G (6 cores, 12 threads) |
| RAM | 16 GB |
| GPU | None (CPU-only) |

## Citation

```bibtex
@misc{axl_2026,
  title={AXL: AXL-Code-1B-Lion - Multi-Scale Transformer for CPU Code Generation},
  author={Koinic},
  year={2026},
  url={https://huggingface.co/KoinicLabs}
}
```

## How to Get Started

### With Ollama

```bash
ollama create axl-code-1b-lion -f Modelfile
ollama run axl-code-1b-lion "def fibonacci():"
```

### With Python

```python
import torch
from multiscale_transformer.model.config import load_config
from multiscale_transformer.model.model import MultiScaleTransformer
from multiscale_transformer.training.tokenizer import ByteTokenizer
config = load_config("config.json")
model = MultiScaleTransformer(config)
ckpt = torch.load("axl_code_1b_lion.pt", map_location="cpu")
model.load_state_dict(ckpt["model_state_dict"])
model.eval()
tokenizer = ByteTokenizer()
prompt = "def fibonacci():"
ids = torch.tensor([tokenizer.encode(prompt)], dtype=torch.long)
with torch.no_grad():
    out = model.generate(ids, max_new_tokens=100, temperature=0.8, top_k=40)
print(tokenizer.decode(out[0].tolist()))
```