README.md

metadata

dataset_name: transformers_code_embeddings
license: apache-2.0
language: code
tags:
  - embeddings
  - transformers-internal
  - similarity-search

Transformers Code Embeddings

Compact index of function/class definitions from src/transformers/models/**/modeling_*.py for cross-model similarity. Built to help surface reusable code when modularizing models.

Contents

• embeddings.safetensors — float32, L2-normalized embeddings shaped [N, D].
• code_index_map.json — {int_id: "relative/path/to/modeling_*.py:SymbolName"}.
• code_index_tokens.json — {identifier: [sorted_unique_tokens]} for Jaccard.

How these were built

• Source: 🤗 Transformers repository, under src/transformers/models.
• Units: top-level class/def definitions.
• Preprocessing:
  • Strip docstrings, comments, and import lines.
  • Replace occurrences of model names and symbol prefixes with Model.
• Encoder: Qwen/Qwen3-Embedding-4B via transformers (mean pooling over tokens, then L2 normalize).
• Output dtype: float32.

Note: Results are tied to a specific Transformers commit. Regenerate when the repo changes.

Quick usage

from huggingface_hub import hf_hub_download
from safetensors.numpy import load_file
import json, numpy as np

repo_id = "hf-internal-testing/transformers_code_embeddings"

emb_path = hf_hub_download(repo_id, "embeddings.safetensors", repo_type="dataset")
map_path = hf_hub_download(repo_id, "code_index_map.json", repo_type="dataset")
tok_path = hf_hub_download(repo_id, "code_index_tokens.json", repo_type="dataset")

emb = load_file(emb_path)["embeddings"]              # (N, D) float32, L2-normalized
id_map = {int(k): v for k, v in json.load(open(map_path))}
tokens = json.load(open(tok_path))

# cosine similarity: dot product
def topk(vec, k=10):
    sims = vec @ emb.T
    idx = np.argpartition(-sims, k)[:k]
    idx = idx[np.argsort(-sims[idx])]
    return [(id_map[int(i)], float(sims[i])) for i in idx]

Intended use

• Identify similar symbols across models (embedding + Jaccard over tokens).
• Assist refactors and modularization efforts.

Limitations

• Embeddings reflect preprocessing choices and the specific encoder.
• Symbols from the same file are present; filter by model name if needed.

Repro/build

See utils/modular_model_detector.py in transformers repo for exact build & push commands.

License

Apache-2.0 for this dataset card and produced artifacts. Source code remains under its original license in the upstream repo.