The architecture is ByteHybrid (3 × Conv1D → 1 × bidirectional attention with RoPE → masked mean-pool → classifier head, with a 4096-bucket trigram-hash embedding), vendored from PleIAs/CommonLingua (Apache-2.0) and trained from scratch on Rosetta Code + The Stack v1 across 107 canonical programming languages.

Comparison with `philomath-1209/programming-language-identification`

3,057 test rows over the 26 labels philomath supports. ONNX, CPUExecutionProvider, batch 64.

model	params	accuracy	macro F1	weighted F1	speed
programming-language-identification-100plus-lite (ONNX)	2.35 M	0.9094	0.9410	0.9361	2.37×
philomath-1209/programming-language-identification (ONNX)	84 M	0.8449	0.8445	0.8467	1.00×

Files

model.pt              fp32 PyTorch checkpoint (CommonLingua format)
model.bf16.pt         bf16 sidecar checkpoint (smaller, same accuracy in eval)
lang2idx.json         107-label index
training_metadata.json  hyperparameters and dataset stats
training_history.json   per-epoch loss / val_acc / val_macro_f1
onnx/
  model.onnx          ONNX export (opset 20, dynamic batch)
  model.onnx.data     external weights blob
  lang2idx.json       (mirror)
  onnx_metadata.json  parity report vs PyTorch

Quick start — PyTorch

python

import torch, numpy as np, sys
sys.path.append("path/to/code-language-id/src")
from code_language_id.byte_hybrid import ByteHybrid, CONFIGS

ckpt = torch.load("model.pt", map_location="cpu", weights_only=False)
model = ByteHybrid(num_classes=ckpt["num_classes"], max_len=ckpt["max_len"],
                   **CONFIGS[ckpt["config"]]).eval()
model.load_state_dict(ckpt["model_state_dict"])
idx2lang = {v: k for k, v in ckpt["lang2idx"].items()}

def encode(texts, max_len=ckpt["max_len"]):
    out = np.full((len(texts), max_len), 256, dtype=np.int64)
    for i, t in enumerate(texts):
        b = t.encode("utf-8", errors="replace")[:max_len]
        out[i, :len(b)] = np.frombuffer(b, dtype=np.uint8)
    return torch.from_numpy(out)

with torch.no_grad():
    logits = model(encode(["def hello():\n    print('hi')"]))
print(idx2lang[int(logits.argmax(-1))])   # -> Python

Quick start — ONNX Runtime

python

import onnxruntime as ort, numpy as np, json

sess = ort.InferenceSession("onnx/model.onnx", providers=["CPUExecutionProvider"])
lang2idx = json.load(open("onnx/lang2idx.json"))
idx2lang = {v: k for k, v in lang2idx.items()}
MAX_LEN = 1023

def encode(texts, max_len=MAX_LEN):
    out = np.full((len(texts), max_len), 256, dtype=np.int64)
    for i, t in enumerate(texts):
        b = t.encode("utf-8", errors="replace")[:max_len]
        out[i, :len(b)] = np.frombuffer(b, dtype=np.uint8)
    return out

logits = sess.run(None, {"byte_ids": encode(["fn main() {}"])})[0]
print(idx2lang[int(logits.argmax(-1))])   # -> Rust

Training summary

Data: Rosetta Code (cakiki/rosetta-code) + The Stack v1 (bigcode/the-stack), task-split to prevent leakage. 72,549 / 9,495 / 8,880 rows (train / val / test) across 107 canonical labels.
Snippets: variable-window (64–1023 bytes) UTF-8.
Optimizer: AdamW (β=0.9, 0.95, weight decay 0.01) + cosine-with-warmup, peak LR 3e-3, 5 % warmup, gradient clipping 1.0.
Schedule: 30 epochs, bf16 autocast, batch 128 (effective 128 with gradient clipping; SDPA fused attention).
Best val macro F1: 0.9085 @ epoch 26 (early stopped).

See training_metadata.json for the full hyperparameter dump.

Citation

If you use this model, please cite:

bibtex

@misc{mariappan2026codelangidlite,
  author    = {Mariappan, Vijayachandran},
  title     = {programming-language-identification-100plus-lite: Byte-level Programming Language Identification across 107 Languages},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/FrameByFrame/programming-language-identification-100plus-lite}
}

Upstream architecture:

bibtex

@misc{commonlingua,
  author    = {{PleIAs}},
  title     = {CommonLingua: Byte-level Language Identification for 334 Languages},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/PleIAs/CommonLingua}
}

License & attribution

Apache-2.0. Architecture and reference inference code derive from PleIAs/CommonLingua (Apache-2.0). Trained weights and dataset curation are original to this repository.

FrameByFrame/programming-language-identification-100plus-lite

programming-language-identification-100plus-lite

Comparison with philomath-1209/programming-language-identification