Datasets:

TinyEdge
/

edge-inference-benchmarks

Error code:   DatasetGenerationCastError
Exception:    DatasetGenerationCastError
Message:      An error occurred while generating the dataset

All the data files must have the same columns, but at some point there are 8 new columns ({'threads', 'decode_tok_s', 'quant', 'prefill_tok_s', 'quality_metric', 'params_b', 'file_mb', 'quality_value'}) and 11 missing columns ({'throughput_per_s', 'accuracy_value', 'eval_set', 'eval_samples', 'accuracy_metric', 'params_m', 'quant_method', 'input_size', 'precision', 'latency_ms_p50', 'batch_size'}).

This happened while the csv dataset builder was generating data using

hf://datasets/TinyEdge/edge-inference-benchmarks/llm-benchmarks.csv (at revision 2d1078b0e69ff8404f38ec75eda841a566b42702), ['hf://datasets/TinyEdge/edge-inference-benchmarks@2d1078b0e69ff8404f38ec75eda841a566b42702/benchmarks.csv', 'hf://datasets/TinyEdge/edge-inference-benchmarks@2d1078b0e69ff8404f38ec75eda841a566b42702/llm-benchmarks.csv']

Please either edit the data files to have matching columns, or separate them into different configurations (see docs at https://hf.co/docs/hub/datasets-manual-configuration#multiple-configurations)
Traceback:    Traceback (most recent call last):
                File "/usr/local/lib/python3.12/site-packages/datasets/builder.py", line 1837, in _prepare_split_single
                  writer.write_table(table)
                File "/usr/local/lib/python3.12/site-packages/datasets/arrow_writer.py", line 765, in write_table
                  self._write_table(pa_table, writer_batch_size=writer_batch_size)
                File "/usr/local/lib/python3.12/site-packages/datasets/arrow_writer.py", line 773, in _write_table
                  pa_table = table_cast(pa_table, self._schema)
                             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.12/site-packages/datasets/table.py", line 2369, in table_cast
                  return cast_table_to_schema(table, schema)
                         ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.12/site-packages/datasets/table.py", line 2297, in cast_table_to_schema
                  raise CastError(
              datasets.table.CastError: Couldn't cast
              snapshot: double
              task: string
              model: string
              params_b: double
              device: string
              device_type: string
              soc: string
              quant: string
              file_mb: double
              runtime: string
              threads: int64
              decode_tok_s: double
              prefill_tok_s: double
              quality_metric: string
              quality_value: double
              sweep: string
              -- schema metadata --
              pandas: '{"index_columns": [{"kind": "range", "name": null, "start": 0, "' + 2118
              to
              {'snapshot': Value('float64'), 'task': Value('string'), 'model': Value('string'), 'params_m': Value('float64'), 'device': Value('string'), 'device_type': Value('string'), 'soc': Value('string'), 'precision': Value('string'), 'quant_method': Value('string'), 'runtime': Value('string'), 'batch_size': Value('int64'), 'input_size': Value('int64'), 'latency_ms_p50': Value('float64'), 'throughput_per_s': Value('float64'), 'accuracy_metric': Value('string'), 'accuracy_value': Value('float64'), 'eval_set': Value('string'), 'eval_samples': Value('int64'), 'sweep': Value('string')}
              because column names don't match
              
              During handling of the above exception, another exception occurred:
              
              Traceback (most recent call last):
                File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 1361, in compute_config_parquet_and_info_response
                  parquet_operations, partial, estimated_dataset_info = stream_convert_to_parquet(
                                                                        ^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 940, in stream_convert_to_parquet
                  builder._prepare_split(split_generator=splits_generators[split], file_format="parquet")
                File "/usr/local/lib/python3.12/site-packages/datasets/builder.py", line 1683, in _prepare_split
                  for job_id, done, content in self._prepare_split_single(
                                               ^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.12/site-packages/datasets/builder.py", line 1839, in _prepare_split_single
                  raise DatasetGenerationCastError.from_cast_error(
              datasets.exceptions.DatasetGenerationCastError: An error occurred while generating the dataset
              
              All the data files must have the same columns, but at some point there are 8 new columns ({'threads', 'decode_tok_s', 'quant', 'prefill_tok_s', 'quality_metric', 'params_b', 'file_mb', 'quality_value'}) and 11 missing columns ({'throughput_per_s', 'accuracy_value', 'eval_set', 'eval_samples', 'accuracy_metric', 'params_m', 'quant_method', 'input_size', 'precision', 'latency_ms_p50', 'batch_size'}).
              
              This happened while the csv dataset builder was generating data using
              
              hf://datasets/TinyEdge/edge-inference-benchmarks/llm-benchmarks.csv (at revision 2d1078b0e69ff8404f38ec75eda841a566b42702), ['hf://datasets/TinyEdge/edge-inference-benchmarks@2d1078b0e69ff8404f38ec75eda841a566b42702/benchmarks.csv', 'hf://datasets/TinyEdge/edge-inference-benchmarks@2d1078b0e69ff8404f38ec75eda841a566b42702/llm-benchmarks.csv']
              
              Please either edit the data files to have matching columns, or separate them into different configurations (see docs at https://hf.co/docs/hub/datasets-manual-configuration#multiple-configurations)

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

snapshot float64	task string	model string	params_m float64	device string	device_type string	soc string	precision string	quant_method string	runtime string	batch_size int64	input_size int64	latency_ms_p50 float64	throughput_per_s float64	accuracy_metric string	accuracy_value float64	eval_set string	eval_samples int64	sweep string
2,026.06	image-classification	mobilenet_v3_small	2.5	oppo-a74	phone	Snapdragon 662	fp32	none	onnxruntime-android 1.20.0 (NNAPI)	1	224	25.99	37.75	imagenetv2_top1	0.536	imagenetv2-mf-500-stratified	500	measured-v2
2,026.06	image-classification	mobilenet_v3_small	2.5	oppo-a74	phone	Snapdragon 662	fp16	nnapi-fp16-flag	onnxruntime-android 1.20.0 (NNAPI fp16)	1	224	26	37.62	imagenetv2_top1	0.536	imagenetv2-mf-500-stratified	500	measured-v2
2,026.06	image-classification	mobilenet_v2	3.5	oppo-a74	phone	Snapdragon 662	fp32	none	onnxruntime-android 1.20.0 (NNAPI)	1	224	62.5	15.84	imagenetv2_top1	0.548	imagenetv2-mf-500-stratified	500	measured-v2
2,026.06	image-classification	mobilenet_v2	3.5	oppo-a74	phone	Snapdragon 662	fp16	nnapi-fp16-flag	onnxruntime-android 1.20.0 (NNAPI fp16)	1	224	62.7	15.54	imagenetv2_top1	0.548	imagenetv2-mf-500-stratified	500	measured-v2
2,026.06	image-classification	mobilenet_v3_large	5.5	oppo-a74	phone	Snapdragon 662	fp32	none	onnxruntime-android 1.20.0 (NNAPI)	1	224	61.59	16.21	imagenetv2_top1	0.576	imagenetv2-mf-500-stratified	500	measured-v2
2,026.06	image-classification	mobilenet_v3_large	5.5	oppo-a74	phone	Snapdragon 662	fp16	nnapi-fp16-flag	onnxruntime-android 1.20.0 (NNAPI fp16)	1	224	61.95	15.9	imagenetv2_top1	0.576	imagenetv2-mf-500-stratified	500	measured-v2
2,026.06	image-classification	resnet18	11.7	oppo-a74	phone	Snapdragon 662	fp32	none	onnxruntime-android 1.20.0 (NNAPI)	1	224	149.74	6.67	imagenetv2_top1	0.552	imagenetv2-mf-500-stratified	500	measured-v2
2,026.06	image-classification	resnet18	11.7	oppo-a74	phone	Snapdragon 662	fp16	nnapi-fp16-flag	onnxruntime-android 1.20.0 (NNAPI fp16)	1	224	150.74	6.6	imagenetv2_top1	0.552	imagenetv2-mf-500-stratified	500	measured-v2
2,026.06	image-classification	efficientnet_b0	5.3	oppo-a74	phone	Snapdragon 662	fp32	none	onnxruntime-android 1.20.0 (NNAPI)	1	224	114.59	8.67	imagenetv2_top1	0.65	imagenetv2-mf-500-stratified	500	measured-v2
2,026.06	image-classification	efficientnet_b0	5.3	oppo-a74	phone	Snapdragon 662	fp16	nnapi-fp16-flag	onnxruntime-android 1.20.0 (NNAPI fp16)	1	224	113.77	8.74	imagenetv2_top1	0.65	imagenetv2-mf-500-stratified	500	measured-v2
2,026.06	image-classification	efficientnet_b2	9.1	oppo-a74	phone	Snapdragon 662	fp32	none	onnxruntime-android 1.20.0 (NNAPI)	1	224	173.16	5.73	imagenetv2_top1	0.642	imagenetv2-mf-500-stratified	500	measured-v2
2,026.06	image-classification	mobilenet_v2	3.5	samsung-sm-p613	tablet	Exynos 9611 (Galaxy Tab S6 Lite 2022)	int8	ort-static-qdq-perchannel	onnxruntime-android 1.20.0 (NNAPI)	1	224	11.92	75.52	imagenetv2_top1	0.548	imagenetv2-mf-500-stratified	500	int8-v1
2,026.06	image-classification	mobilenet_v2	3.5	oppo-a74	phone	Snapdragon 662	int8	ort-static-qdq-perchannel	onnxruntime-android 1.20.0 (NNAPI)	1	224	22.54	42.87	imagenetv2_top1	0.548	imagenetv2-mf-500-stratified	500	int8-v1
2,026.06	image-classification	mobilenet_v3_large	5.5	samsung-sm-p613	tablet	Exynos 9611 (Galaxy Tab S6 Lite 2022)	int8	ort-static-qdq-perchannel	onnxruntime-android 1.20.0 (NNAPI)	1	224	14.46	60.88	imagenetv2_top1	0.536	imagenetv2-mf-500-stratified	500	int8-v1
2,026.06	image-classification	mobilenet_v3_large	5.5	oppo-a74	phone	Snapdragon 662	int8	ort-static-qdq-perchannel	onnxruntime-android 1.20.0 (NNAPI)	1	224	29.22	33.95	imagenetv2_top1	0.536	imagenetv2-mf-500-stratified	500	int8-v1
2,026.06	text-generation	qwen2.5-0.5b-instruct	null	oppo-a74	phone	Snapdragon 662	null	null	llama.cpp b9590 (Termux, CPU)	null	null	null	null	null	null	null	null	llm-quant-v1
2,026.06	text-generation	qwen2.5-0.5b-instruct	null	oppo-a74	phone	Snapdragon 662	null	null	llama.cpp b9590 (Termux, CPU)	null	null	null	null	null	null	null	null	llm-quant-v1
2,026.06	text-generation	qwen2.5-0.5b-instruct	null	oppo-a74	phone	Snapdragon 662	null	null	llama.cpp b9590 (Termux, CPU)	null	null	null	null	null	null	null	null	llm-quant-v1
2,026.06	text-generation	qwen2.5-0.5b-instruct	null	oppo-a74	phone	Snapdragon 662	null	null	llama.cpp b9590 (Termux, CPU)	null	null	null	null	null	null	null	null	llm-quant-v1
2,026.06	text-generation	qwen2.5-0.5b-instruct	null	oppo-a74	phone	Snapdragon 662	null	null	llama.cpp b9590 (Termux, CPU)	null	null	null	null	null	null	null	null	llm-quant-v1

TinyEdge edge-inference benchmarks

Independently measured latency and accuracy for well-known vision models on real edge devices (phones, tablets — fleet growing), produced by TinyEdge, a device cloud for edge-AI benchmarking. Nothing here is taken from papers or spec sheets: every row is a job executed on the physical device through TinyEdge's production agent, with accuracy measured on a fixed 500-image stratified sample of ImageNet-V2 (matched-frequency) using a standardized preprocessing spec, batch 1, 224x224.

ImageNet-V2 top-1 runs systematically below classic ImageNet val for every model (known distribution shift) — treat accuracy_value as a relative capacity indicator; rankings and deltas are the signal.

Headline findings (snapshot 2026.06)

fp16 is a placebo on these NPUs (onnxruntime-android + NNAPI): identical latency and accuracy to fp32 on every model tested.
Static int8 (QDQ) is the real lever: 2.6–2.8x faster at bit-identical accuracy for MobileNetV2 — when the architecture tolerates it.
Quantization safety is architecture-dependent: naive post-training int8 collapsed MobileNetV3-Small (−57 pts) and EfficientNet-B0 (−32 pts); MobileNetV3-Large lost 4 pts; MobileNetV2 lost nothing.
Dynamic ("data-free") int8 does not run at all on Android ONNX Runtime (ConvInteger not implemented).
Published accuracy can invert on-device rankings: EfficientNet-B2 (published +2.9 over B0) measured no better than B0 while 51% slower.

LLM benchmarks (llm-benchmarks.csv)

GGUF quantization ladders measured on-device via llama.cpp (CPU): decode and prefill tok/s at the measured-optimal thread count (big cores only — the llama.cpp default of all-cores is 28-39% slower on big.LITTLE SoCs), with WikiText-2 perplexity (fixed 32-chunk subset) as the device-independent quality column. First snapshot: Qwen2.5-0.5B-Instruct, Q8_0→Q3_K_M, on a Snapdragon 662 phone — 14-19 tok/s decode (conversational) at every level; Q6_K dominates Q8_0; prefill (~30-40 tok/s) is the practical bottleneck for long prompts.

Schema

One row = one (model, device, precision/quant) measurement. Key columns: latency_ms_p50, throughput_per_s, accuracy_value (accuracy_metric = imagenetv2_top1), quant_method, runtime (includes the execution provider that actually ran), sweep (provenance). The task column is image-classification for all current rows; LLM rows (decode tok/s, TTFT, RAM) will use the same table with task=text-generation.

Method & provenance

Sweep tooling and raw per-run JSON live in the TinyEdge repos; results are snapshot-versioned via the snapshot column. Jobs run on battery-policy-aware agents; each result records the exact runtime + execution provider. Eval images are NOT redistributed here — only measurements.

Questions / want your model or device measured: https://tinyedge.ai/contact

Downloads last month: 19

Total file size:

10.3 kB