Datasets:

gllllll
/

glsl-opengl-educational-dataset

license: other
task_categories:
  - text-generation
language:
  - en
tags:
  - graphics
  - glsl
  - wgsl
  - hlsl
  - metal
  - spir-v
  - webgpu
  - shaders
  - compiler-validated
size_categories:
  - 10K<n<100K

ShaderCross Graphics Dataset v1.1

Multi-dialect shader corpus where every dialect is produced by a real compiler and validated by a real validator.

What changed from the previous release

The earlier release advertised a five-way polyglot matrix but generated it with regular-expression substitution. Measured on the published files:

Claim	Measured reality
Five dialects via compilers	`wgsl_transpiler.py` was ~60 regexes; `subprocess` was imported but never called
Valid targets	`dxc -T ps_6_0` rejected the HLSL: `layout(location = 0) in float3 v_WorldPos;`
Distinct implementations	`quick_ratio(glsl, wgsl) = 0.9996`; 82% of records were four near-identical copies
Zero-junk filtering	`validation` was all zeros; 42% of records were not shaders (e.g. a tree-sitter C scanner)
Dense corpus	41.6% of bytes were exact duplicates stored under two keys

v1.1 replaces the transpiler with glslang -> spirv-cross -> naga and validates each target with spirv-val, glslangValidator, dxc and wgpu. On the same GGX shader, GLSL-to-HLSL similarity fell from 0.9996 to 0.715 and dxc now accepts the output.

Generation and validation matrix

Dialect	Generator	This run	Validator
SPIR-V	glslangValidator -V	855/860 (99%)	spirv-val
GLSL 330	spirv-cross	709/829 (86%)	glslangValidator
ESSL 300	spirv-cross	695/752 (92%)	glslangValidator
HLSL SM6.0	spirv-cross	708/784 (90%)	dxc (authoritative)
MSL 2.0	spirv-cross	emitted, unverified	none exists on Linux
WGSL	sampler-split + naga	544/719 (76%)	wgpu.create_shader_module

msl_status is reported as EMITTED_UNVERIFIED, never PASSED: no Metal compiler exists outside macOS, and overstating that is what made the previous release untrustworthy.

The WGSL fix

GLSL's sampler2D is a combined image sampler; WGSL requires separate texture_2d<f32> and sampler bindings. naga cannot bridge that gap, so 64% of shaders failed conversion with invalid id %N. Twelve SPIR-V hygiene strategies scored 0/40. Normalising the source instead lifted WGSL from 36% to 86%.

File	Records
`dataset_full.jsonl`	566
`dataset_permissive.jsonl`	537
`dataset_noncommercial.jsonl`	29
`diagnostic_bugfix_pairs.jsonl`	2076
`metrics.json`	pipeline counters

Parquet mirrors use nested Arrow structs rather than JSON-in-a-string.

Diagnostic pairs

Compilation failures are not discarded. Each becomes a record with the compiler invocation, its exact log, and an explanation. Cases marked is_fixable: false are constructs that cannot exist in the target API at all - subpass input, gl_PointCoord, mesh shaders, descriptor indexing - which teaches genuine portability boundaries.

Pipeline statistics for this build

repositories : 64/102 cloned
files scanned: 5769
candidates   : 2893
accepted     : 566
diagnostics  : 2076
rejected     : L0=2579 licence=297 inflation=23
duplicates   : exact=90 near=144
tokens       : 3,871,248
elapsed      : 673s

Toolchain

{
  "glslang": "Glslang Version: 10:11.8.0",
  "spirv_cross": "(Debian package)",
  "spirv_val": "SPIRV-Tools v2022.2-dev unknown hash, 2022-02-16T16:37:15",
  "spirv_opt": "SPIRV-Tools v2022.2-dev unknown hash, 2022-02-16T16:37:15",
  "spirv_dis": "SPIRV-Tools v2022.2-dev unknown hash, 2022-02-16T16:37:15",
  "naga": "30.0.0",
  "dxc": "libdxcompiler.so: 1.8(dev;3944-3e105849)"
}

Licensing

Records carry license and license_category. Use dataset_permissive.jsonl for commercial pre-training; dataset_noncommercial.jsonl holds CC-BY-NC and GPL-derived material for research only.

Schema

See docs/ARCHITECTURE_v1.1.md in the source repository. Key point: generated code lives in exactly one place, implementations.*, and validation is tri-state (PASSED / FAILED / EMITTED_UNVERIFIED) so unprovable claims are representable rather than rounded up to success.