TypeGround: A Compiler-Grounded Benchmark for TypeScript Type Inference

TypeGround is a compiler-grounded benchmark for TypeScript type inference. Every stage — annotation, context construction, and quality control — is validated by the TypeScript Compiler (TSC). It extracts variable declarations, function parameters, and return types from compilable TypeScript projects, combining developer annotations, TSC inference, LLM candidate generation, and TSC backfill checking to produce type labels.

Covers 2,735 projects, 99,209 source files, and 1,428,997 type annotations.

Benchmark Format

Each sample targets a single type position and includes:

• Metadata (name, type, loc_cat, type_cat, line, file)
• Snippet code: ±25 lines around the target, preserving local lexical and syntactic context.
• Slice code: type-dependency-based statements collected across function boundaries (imports, definitions, call sites, use sites).

The target type is replaced by <mask>; the model predicts the masked type expression.

Location Categories

Value	Meaning
var	Variable declaration
arg	Function parameter
ret	Function return

Type Categories

The type_cat field stores granular TSC type classifications (e.g., NumberKeyword, FunctionType, UnionType, userDefined). type_cat.py maps them into three analysis categories:

Category	Coverage
BuiltIn	TS keywords (*Keyword), built-in objects (Date, Error, …), arrays/buffers, DOM (HTML*, SVG*, …), events (*Event), Canvas/WebGL, browser/Web APIs
Constructed	Function/constructor types, composite types (ArrayType, UnionType, IntersectionType, …), utility types (Partial, Pick, Omit, …), type operators (MappedType, ConditionalType, …)
UserDefined	userDefined — project-declared interfaces, type aliases, classes, enums

Data Files

File	Records	Size	Description
types.json.gz	~11.4M	21 MB	Train metadata
types_testdata.json.gz	~694K	1.3 MB	Test metadata
snippet_data_50.jsonl.gz	~802K	203 MB	Train snippets
snippet_data_testdata_50.jsonl.gz	~41K	11 MB	Test snippets
slice_data.jsonl.gz	~1.2M	307 MB	Train slices
slice_data_testdata.jsonl.gz	~58K	15 MB	Test slices

Format

Metadata (types.json.gz):

{"name":"smooth","type":"number","loc_cat":"ret","type_cat":"NumberKeyword","line":1,"file":"repo/path/to/file.ts"}

Snippet / Slice (snippet_data_50.jsonl.gz / slice_data.jsonl.gz, one JSON object per line):

{"id":"repo/path/to/file.ts#func#var#2","maskedText":"... <mask> ...","type":"number"}
{"id":"repo/path/to/file.ts#func#var#2","sliced_code":"... <mask> ...","type":"number"}

Configs

Config	Train file	Test file	Columns
snippet	snippet_data_50.jsonl.gz	snippet_data_testdata_50.jsonl.gz	id, maskedText, type
slice	slice_data.jsonl.gz	slice_data_testdata.jsonl.gz	id, sliced_code, type

Usage

from datasets import load_dataset

# Snippet context
snippet = load_dataset("fumx66/TypeGround", "snippet")
print(snippet["train"][0]["maskedText"])

# Slice context
slice_ds = load_dataset("fumx66/TypeGround", "slice")
print(slice_ds["train"][0]["sliced_code"])

# Metadata
import gzip, json
with gzip.open("data/types.json.gz", "rt") as f:
    types = json.load(f)

Citation

@dataset{typeground,
  title={TypeGround: A Compiler-Grounded Benchmark for TypeScript Type Inference},
  year={2026},
  url={https://huggingface.co/datasets/fumx66/TypeGround}
}

License

MIT License