53 models · 6 datasets · 2 collections · 7,206 total downloads
MLX artifacts for the IEEE Cloud Summit 2026 paper. 3 LLMs × 5 precisions (BF16/Q8/Q6/Q4/Q3, gs=64).
36 pruned-LLM artifacts (3 models x 3 methods x 4 sparsities) for AIIoT 2026 bias-amplification paper. Research only - not for deployment.
StableHLO-OutOfGrammar-25 Hand-authored NL→StableHLO pairs using ops outside our 10-op grammar scope (n=25, stress set). Composition Instances: 25 Format: one JSON record per line in data/test.jsonl Schema: fields = dialect, id, mlir, nl, op, source Verifier: stablehlo-opt v1.4.0 (upstream truth) and iree-compile --compile-to=input (substitute, 50/50 concordant on a stratified n=50 sample) License: Apache-2.0 (SPDX: Apache-2.0). No third-party IP restrictions.… See the full description on the dataset page: https://huggingface.co/datasets/plawanrath/StableHLO-OutOfGrammar-25.
Linalg-Spec-30 Hand-authored NL→MLIR pairs for linalg named ops under memref semantics (n=30). Composition Instances: 30 Format: one JSON record per line in data/test.jsonl Schema: fields = dialect, difficulty, id, mlir, nl, notes Verifier: mlir-opt --verify-diagnostics against pinned LLVM 19.1.7 License: Apache-2.0 (SPDX: Apache-2.0). No third-party IP restrictions. Loading from datasets import load_dataset ds = load_dataset("plawanrath/Linalg-Spec-30"… See the full description on the dataset page: https://huggingface.co/datasets/plawanrath/Linalg-Spec-30.
StableHLO-Spec-30 Hand-authored NL→StableHLO pairs across 10 op families (n=30). Composition Instances: 30 Format: one JSON record per line in data/test.jsonl Schema: fields = dialect, difficulty, id, mlir, nl, notes Verifier: stablehlo-opt v1.4.0 (upstream truth) and iree-compile --compile-to=input (substitute, 50/50 concordant on a stratified n=50 sample) License: Apache-2.0 (SPDX: Apache-2.0). No third-party IP restrictions. Loading from datasets… See the full description on the dataset page: https://huggingface.co/datasets/plawanrath/StableHLO-Spec-30.
StableHLO-Held-Out-200 Programmatic parametric sweep over 7 StableHLO op families × 6 dtypes × 3 shape ranks (n=200, verifier-clean). Composition Instances: 200 Format: one JSON record per line in data/test.jsonl Schema: fields = dialect, difficulty, id, mlir, nl, notes, source Verifier: stablehlo-opt v1.4.0 (upstream truth) and iree-compile --compile-to=input (substitute, 50/50 concordant on a stratified n=50 sample) License: Apache-2.0 (SPDX: Apache-2.0). No… See the full description on the dataset page: https://huggingface.co/datasets/plawanrath/StableHLO-Held-Out-200.
MLIR-Spec-150 Hand-authored NL→MLIR pairs for arith+func+memref dialects (n=150). Composition Instances: 150 Format: one JSON record per line in data/test.jsonl Schema: fields = dialect, difficulty, id, mlir, nl, notes Verifier: mlir-opt --verify-diagnostics against pinned LLVM 19.1.7 License: Apache-2.0 (SPDX: Apache-2.0). No third-party IP restrictions. Loading from datasets import load_dataset ds = load_dataset("plawanrath/MLIR-Spec-150", split="test")… See the full description on the dataset page: https://huggingface.co/datasets/plawanrath/MLIR-Spec-150.
MLIR-Functional-Reference-30 Hand-authored functional-correctness reference set for arith, linalg+memref, and stablehlo (n=30, 10 per dialect). Composition Instances: 30 Format: one JSON record per line in data/test.jsonl Schema: fields = canonical_fn_name, canonical_signature, dialect, expected_output, expected_output_pattern, expected_stdout_regex, id, inputs, iree_inputs, memref_inputs, memref_print, nl, result_type, scalar_inputs, source_benchmark, source_id… See the full description on the dataset page: https://huggingface.co/datasets/plawanrath/MLIR-Functional-Reference-30.