Dataset Viewer
The dataset viewer is not available for this split.
Cannot load the dataset split (in streaming mode) to extract the first rows.
Error code: StreamingRowsError
Exception: CastError
Message: Couldn't cast
device: string
grep: string
head: string
per_image: list<item: struct<idx: int64, ref: string, pred: string, new_tokens: int64, gen_s: double, wall_s: d (... 97 chars omitted)
child 0, item: struct<idx: int64, ref: string, pred: string, new_tokens: int64, gen_s: double, wall_s: double, pref (... 85 chars omitted)
child 0, idx: int64
child 1, ref: string
child 2, pred: string
child 3, new_tokens: int64
child 4, gen_s: double
child 5, wall_s: double
child 6, prefill_ms: double
child 7, prompt_per_second: double
child 8, predicted_per_second: double
child 9, cer: double
n_images: int64
e2e_decode_tps: double
variant: string
total_gen_s: double
llama_cpp_note: string
prefill_s_total: double
model: string
total_tokens: int64
mean_cer: double
prompt: string
to
{'model': Value('string'), 'variant': Value('string'), 'device': Value('string'), 'n_images': Value('int64'), 'prompt': Value('string'), 'e2e_decode_tps': Value('float64'), 'mean_cer': Value('float64'), 'total_gen_s': Value('float64'), 'total_tokens': Value('int64'), 'prefill_s_total': Value('float64'), 'llama_cpp_note': Value('string'), 'per_image': List({'idx': Value('int64'), 'ref': Value('string'), 'pred': Value('string'), 'new_tokens': Value('int64'), 'gen_s': Value('float64'), 'wall_s': Value('float64'), 'prefill_ms': Value('float64'), 'prompt_per_second': Value('float64'), 'predicted_per_second': Value('float64'), 'cer': Value('float64')})}
because column names don't match
Traceback: Traceback (most recent call last):
File "/src/services/worker/src/worker/utils.py", line 149, in get_rows_or_raise
return get_rows(
dataset=dataset,
...<4 lines>...
column_names=column_names,
)
File "/src/libs/libcommon/src/libcommon/utils.py", line 272, in decorator
return func(*args, **kwargs)
File "/src/services/worker/src/worker/utils.py", line 129, in get_rows
rows_plus_one = list(itertools.islice(safe_iter(ds, dataset=dataset), rows_max_number + 1))
File "/src/services/worker/src/worker/utils.py", line 544, in safe_iter
yield from ds.decode(False) if ds.features else ds
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 3052, in __iter__
for key, example in ex_iterable:
^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2559, in __iter__
for key, pa_table in self._iter_arrow():
~~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2584, in _iter_arrow
for key, pa_table in self.ex_iterable._iter_arrow():
~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 558, in _iter_arrow
for key, pa_table in iterator:
^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 440, in _iter_arrow
for key, pa_table in self.generate_tables_fn(**gen_kwags):
~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 361, in _generate_tables
self._cast_table(pa_table, json_field_paths=json_field_paths),
~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 132, in _cast_table
pa_table = table_cast(pa_table, self.info.features.arrow_schema)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2381, in table_cast
return cast_table_to_schema(table, schema)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2309, in cast_table_to_schema
raise CastError(
...<3 lines>...
)
datasets.table.CastError: Couldn't cast
device: string
grep: string
head: string
per_image: list<item: struct<idx: int64, ref: string, pred: string, new_tokens: int64, gen_s: double, wall_s: d (... 97 chars omitted)
child 0, item: struct<idx: int64, ref: string, pred: string, new_tokens: int64, gen_s: double, wall_s: double, pref (... 85 chars omitted)
child 0, idx: int64
child 1, ref: string
child 2, pred: string
child 3, new_tokens: int64
child 4, gen_s: double
child 5, wall_s: double
child 6, prefill_ms: double
child 7, prompt_per_second: double
child 8, predicted_per_second: double
child 9, cer: double
n_images: int64
e2e_decode_tps: double
variant: string
total_gen_s: double
llama_cpp_note: string
prefill_s_total: double
model: string
total_tokens: int64
mean_cer: double
prompt: string
to
{'model': Value('string'), 'variant': Value('string'), 'device': Value('string'), 'n_images': Value('int64'), 'prompt': Value('string'), 'e2e_decode_tps': Value('float64'), 'mean_cer': Value('float64'), 'total_gen_s': Value('float64'), 'total_tokens': Value('int64'), 'prefill_s_total': Value('float64'), 'llama_cpp_note': Value('string'), 'per_image': List({'idx': Value('int64'), 'ref': Value('string'), 'pred': Value('string'), 'new_tokens': Value('int64'), 'gen_s': Value('float64'), 'wall_s': Value('float64'), 'prefill_ms': Value('float64'), 'prompt_per_second': Value('float64'), 'predicted_per_second': Value('float64'), 'cer': Value('float64')})}
because column names don't matchNeed help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
MiniCPM-OCR-1.0 quantization: which version on which hardware
One-table guide from an measured benchmark of every official quantization of openbmb/MiniCPM-OCR-1.0 (BF16, GGUF Q4_K_M, AWQ 4-bit, GPTQ 4-bit) on an NVIDIA T4 and on 8 vCPU, over 20 OCR images with ground truth.
The table
| Hardware | Use this | Why | Avoid | Numbers (s/image · CER) |
|---|---|---|---|---|
| T4 / any pre-Ampere GPU (SM 7.5) | GGUF Q4_K_M via llama.cpp | Fastest end-to-end, 2× smaller than the next option, ~lossless. FP16 BF16-repo is a close second if you want HF-native. | GPTQ: only runs on the slow pure-torch kernel here (no prebuilt SM 7.5 kernel). | GGUF 0.47 s · CER 0.010 · BF16-fp16 0.59 s · 0.007 · AWQ 1.06 s · 0.007 · GPTQ 1.25 s · 0.011 |
| T4, HF-native (transformers) required | BF16 repo (as fp16) | Fastest transformers leg and the accuracy reference. AWQ loses 45% throughput for zero accuracy gain. | GPTQ (kernel fallback, see above). | BF16-fp16 0.59 s · 0.007 |
| 8-vCPU CPU server / laptop | GGUF Q4_K_M via llama.cpp | 3× faster than fp32 transformers, ~1.4 GB total (529 MB LLM + 901 MB mmproj). | AWQ (62 s/image — 4× slower than fp32!) and GPTQ (crashes natively). | GGUF 5.2 s · CER 0.010 · BF16-fp32 15.8 s · 0.007 |
| Low-RAM CPU (2–4 GB) | GGUF Q4_K_M | The only variant that fits: 529 MB LLM. The vision encoder (mmproj, fp16) dominates both memory and CPU prefill. | Everything safetensors-based (≥1.9 GB). | GGUF: decode 51 tok/s, 8 threads |
Accuracy (CER, lower is better — 20 SROIE text-line images)
All quants are effectively lossless on OCR text lines: 0.007 (BF16) ≈ 0.007 (AWQ) < 0.0095 (GGUF Q4_K_M) < 0.011 (GPTQ). Per-image outputs are near-identical across variants (same 3 characters wrong in the same 3 images for BF16/AWQ/GGUF). Quantize freely; pick by speed and memory, not accuracy.
Details that matter
- T4 has no BF16: the BF16 repo runs as fp16 there; on CPU as fp32.
- GPTQ on T4 needed
GPTQConfig(bits=4, backend="gptq_torch")— transformers' auto backend tries gptqmodel's Marlin kernel, which needs SM ≥ 8.0 or a JIT build (nvcc) that the container lacks. The torch fallback is what the 1.25 s/image measures; a properly-kernel'd GPTQ would likely beat AWQ on SM ≥ 8.0 GPUs. - GPTQ on CPU does not work with transformers 5.22.1 + gptqmodel 7.7.0: the process SIGABRTs inside
TorchLinearcompilation (reproduced 4×). - GGUF memory is dominated by the vision encoder: mmproj-model-f16-4x.gguf is 901 MB vs 529 MB for Q4_K_M — 1.77 GiB VRAM measured for the whole server on T4.
- llama.cpp on HF GPU jobs: the prebuilt
ubuntu-cuda-*tarball silently falls back to CPU in containers without the CUDA runtime wheels; the officialghcr.io/ggml-org/llama.cpp:server-cudaimage works and gives ~206 tok/s decode, ~730 tok/s prefill (first request adds ~59 s CUDA-graph warmup). - Speed metric: end-to-end seconds per image (prompt + generation, ~9 output tokens avg on this eval), identical prompt (
Text Recognition:), greedy decoding, batch size 1, same 20 images for every cell.
Raw data
results/<hardware>_<variant>.json— per-image predictions, timings, CER for every cell (the source of this table)results/cpu8_tf_gptq_ERROR.json— the CPU GPTQ failure evidencesummary.csv— machine-readable version of the tablescripts/— the benchmark harness (transformers + llama.cpp drivers)
Method: transformers==5.22.1 (AutoModelForImageTextToText), gptqmodel 7.7.0, llama.cpp b12014, eval set = 20 rows of priyank-m/SROIE_2019_text_recognition (test), CER via jiwer after whitespace/case normalization. Benchmark produced by ML Intern on 2026-10-30.
- Downloads last month
- 340