Hugging Face's logo

Datasets:
lazybrick
/
kiln-evals

Size:
10K<n<100K
Tags:
evaluation
quantization
qwen3.5
lm-evaluation-harness
vlmevalkit
License:
Dataset card Files
xet
 Community
Dataset Viewer
Duplicate
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
results: struct<mmlu_pro_biology: struct<name: string, alias: string, sample_len: int64, exact_match,custom-e (... 3653 chars omitted)
  child 0, mmlu_pro_biology: struct<name: string, alias: string, sample_len: int64, exact_match,custom-extract: double, exact_mat (... 33 chars omitted)
      child 0, name: string
      child 1, alias: string
      child 2, sample_len: int64
      child 3, exact_match,custom-extract: double
      child 4, exact_match_stderr,custom-extract: double
  child 1, mmlu_pro_business: struct<name: string, alias: string, sample_len: int64, exact_match,custom-extract: double, exact_mat (... 33 chars omitted)
      child 0, name: string
      child 1, alias: string
      child 2, sample_len: int64
      child 3, exact_match,custom-extract: double
      child 4, exact_match_stderr,custom-extract: double
  child 2, mmlu_pro_chemistry: struct<name: string, alias: string, sample_len: int64, exact_match,custom-extract: double, exact_mat (... 33 chars omitted)
      child 0, name: string
      child 1, alias: string
      child 2, sample_len: int64
      child 3, exact_match,custom-extract: double
      child 4, exact_match_stderr,custom-extract: double
  child 3, mmlu_pro_computer_science: struct<name: string, alias: string, sample_len: int64, exact_match,custom-extract: double, exact_mat (... 33 chars omitted)
      child 0, name: string
      child 1, alias: string
      child 2, sample_len: int64
      child 3, exact_match,custom-extract: double
      chil
...
tions: struct<MMBench_DEV_EN_V11: struct<file: string, sha256: string>, MMMU_DEV_VAL: struct<file: string,  (... 70 chars omitted)
      child 0, MMBench_DEV_EN_V11: struct<file: string, sha256: string>
          child 0, file: string
          child 1, sha256: string
      child 1, MMMU_DEV_VAL: struct<file: string, sha256: string>
          child 0, file: string
          child 1, sha256: string
      child 2, MathVista_MINI: struct<file: string, sha256: string>
          child 0, file: string
          child 1, sha256: string
  child 4, extraction_failures: struct<Qwen3.5-4B-pinned_MMBench_DEV_EN_V11_gpt-4o-mini_result.xlsx: int64, Qwen3.5-4B-pinned_MMMU_D (... 96 chars omitted)
      child 0, Qwen3.5-4B-pinned_MMBench_DEV_EN_V11_gpt-4o-mini_result.xlsx: int64
      child 1, Qwen3.5-4B-pinned_MMMU_DEV_VAL_gpt-4o-mini_result.xlsx: int64
      child 2, Qwen3.5-4B-pinned_MathVista_MINI_gpt-4o-mini.xlsx: int64
  child 5, final_answer_rule: struct<applies: string, answers: struct<Qwen3.5-4B-pinned_MMBench_DEV_EN_V11_gpt-4o-mini_result.xlsx (... 130 chars omitted)
      child 0, applies: string
      child 1, answers: struct<Qwen3.5-4B-pinned_MMBench_DEV_EN_V11_gpt-4o-mini_result.xlsx: int64, Qwen3.5-4B-pinned_MMMU_D (... 96 chars omitted)
          child 0, Qwen3.5-4B-pinned_MMBench_DEV_EN_V11_gpt-4o-mini_result.xlsx: int64
          child 1, Qwen3.5-4B-pinned_MMMU_DEV_VAL_gpt-4o-mini_result.xlsx: int64
          child 2, Qwen3.5-4B-pinned_MathVista_MINI_gpt-4o-mini.xlsx: int64
to
{'model': Value('string'), 'runs': {'suite': {'note': Value('string'), 'benchmark_suite': {'enable_thinking': Value('bool'), 'max_gen_toks': Value('int64'), 'max_model_len': Value('int64'), 'path': Value('string'), 'samples': Value('null'), 'seeds': List(Value('int64')), 'sha256': Value('string')}, 'code_revision': Value('string'), 'config_digest': Value('string'), 'durations_seconds': {}, 'enable_thinking': Value('bool'), 'feature': Value('string'), 'from_hub': Value('string'), 'limit': Value('null'), 'method': Value('string'), 'model_id': Value('string'), 'model_revision': Value('string'), 'research_result': Value('bool'), 'schema_version': Value('int64'), 'scope': Value('string'), 'status': Value('string'), 'variant': Value('string'), 'vision_limit': Value('null'), 'vision_protocol': {'enable_thinking': Value('bool'), 'inference': Value('string'), 'judge': {'MMBench_DEV_EN_V11': Value('string'), 'MMMU_DEV_VAL': Value('string'), 'MathVista_MINI': Value('string'), 'default': Value('string')}, 'max_model_len': Value('int64'), 'max_tokens': Value('int64'), 'temperature': Value('float64')}}, 'docvqa': {'note': Value('string'), 'benchmark_suite': {'enable_thinking': Value('bool'), 'max_gen_toks': Value('int64'), 'max_model_len': Value('int64'), 'path': Value('string'), 'samples': Value('null'), 'seeds': List(Value('int64')), 'sha256': Value('string')}, 'code_revision': Value('string'), 'config_digest': Value('string'), 'durations_seconds': {'runtime_smoke': Value('float64'), 'vi
...
'), 'model_id': Value('string'), 'model_revision': Value('string'), 'research_result': Value('bool'), 'schema_version': Value('int64'), 'scope': Value('string'), 'status': Value('string'), 'variant': Value('string'), 'vision_limit': Value('null'), 'vision_protocol': {'enable_thinking': Value('bool'), 'inference': Value('string'), 'judge': {'default': Value('string')}, 'max_model_len': Value('int64'), 'max_tokens': Value('int64'), 'temperature': Value('float64')}}}, 'vision_rule_scoring': {'tasks': List(Value('string')), 'note': Value('string'), 'vlmevalkit_revision': Value('string')}, 'vision_answer_extraction': {'answer_extractor': Value('string'), 'tasks': List(Value('string')), 'vlmevalkit_revision': Value('string'), 'predictions': {'MMBench_DEV_EN_V11': {'file': Value('string'), 'sha256': Value('string')}, 'MMMU_DEV_VAL': {'file': Value('string'), 'sha256': Value('string')}, 'MathVista_MINI': {'file': Value('string'), 'sha256': Value('string')}}, 'extraction_failures': {'Qwen3.5-4B-pinned_MMBench_DEV_EN_V11_gpt-4o-mini_result.xlsx': Value('int64'), 'Qwen3.5-4B-pinned_MMMU_DEV_VAL_gpt-4o-mini_result.xlsx': Value('int64'), 'Qwen3.5-4B-pinned_MathVista_MINI_gpt-4o-mini.xlsx': Value('int64')}, 'final_answer_rule': {'applies': Value('string'), 'answers': {'Qwen3.5-4B-pinned_MMBench_DEV_EN_V11_gpt-4o-mini_result.xlsx': Value('int64'), 'Qwen3.5-4B-pinned_MMMU_DEV_VAL_gpt-4o-mini_result.xlsx': Value('int64'), 'Qwen3.5-4B-pinned_MathVista_MINI_gpt-4o-mini.xlsx': Value('int64')}}}}
because column names don't match
Traceback:    Traceback (most recent call last):
                File "/src/services/worker/src/worker/utils.py", line 147, 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 127, 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 483, 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 2951, in __iter__
                  for key, example in ex_iterable:
                                      ^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2461, in __iter__
                  for key, pa_table in self._iter_arrow():
                                       ~~~~~~~~~~~~~~~~^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2486, 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 547, in _iter_arrow
                  for key, pa_table in iterator:
                                       ^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 430, 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 343, 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 2378, in table_cast
                  return cast_table_to_schema(table, schema)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2306, in cast_table_to_schema
                  raise CastError(
                  ...<3 lines>...
                  )
              datasets.table.CastError: Couldn't cast
              results: struct<mmlu_pro_biology: struct<name: string, alias: string, sample_len: int64, exact_match,custom-e (... 3653 chars omitted)
                child 0, mmlu_pro_biology: struct<name: string, alias: string, sample_len: int64, exact_match,custom-extract: double, exact_mat (... 33 chars omitted)
                    child 0, name: string
                    child 1, alias: string
                    child 2, sample_len: int64
                    child 3, exact_match,custom-extract: double
                    child 4, exact_match_stderr,custom-extract: double
                child 1, mmlu_pro_business: struct<name: string, alias: string, sample_len: int64, exact_match,custom-extract: double, exact_mat (... 33 chars omitted)
                    child 0, name: string
                    child 1, alias: string
                    child 2, sample_len: int64
                    child 3, exact_match,custom-extract: double
                    child 4, exact_match_stderr,custom-extract: double
                child 2, mmlu_pro_chemistry: struct<name: string, alias: string, sample_len: int64, exact_match,custom-extract: double, exact_mat (... 33 chars omitted)
                    child 0, name: string
                    child 1, alias: string
                    child 2, sample_len: int64
                    child 3, exact_match,custom-extract: double
                    child 4, exact_match_stderr,custom-extract: double
                child 3, mmlu_pro_computer_science: struct<name: string, alias: string, sample_len: int64, exact_match,custom-extract: double, exact_mat (... 33 chars omitted)
                    child 0, name: string
                    child 1, alias: string
                    child 2, sample_len: int64
                    child 3, exact_match,custom-extract: double
                    chil
              ...
              tions: struct<MMBench_DEV_EN_V11: struct<file: string, sha256: string>, MMMU_DEV_VAL: struct<file: string,  (... 70 chars omitted)
                    child 0, MMBench_DEV_EN_V11: struct<file: string, sha256: string>
                        child 0, file: string
                        child 1, sha256: string
                    child 1, MMMU_DEV_VAL: struct<file: string, sha256: string>
                        child 0, file: string
                        child 1, sha256: string
                    child 2, MathVista_MINI: struct<file: string, sha256: string>
                        child 0, file: string
                        child 1, sha256: string
                child 4, extraction_failures: struct<Qwen3.5-4B-pinned_MMBench_DEV_EN_V11_gpt-4o-mini_result.xlsx: int64, Qwen3.5-4B-pinned_MMMU_D (... 96 chars omitted)
                    child 0, Qwen3.5-4B-pinned_MMBench_DEV_EN_V11_gpt-4o-mini_result.xlsx: int64
                    child 1, Qwen3.5-4B-pinned_MMMU_DEV_VAL_gpt-4o-mini_result.xlsx: int64
                    child 2, Qwen3.5-4B-pinned_MathVista_MINI_gpt-4o-mini.xlsx: int64
                child 5, final_answer_rule: struct<applies: string, answers: struct<Qwen3.5-4B-pinned_MMBench_DEV_EN_V11_gpt-4o-mini_result.xlsx (... 130 chars omitted)
                    child 0, applies: string
                    child 1, answers: struct<Qwen3.5-4B-pinned_MMBench_DEV_EN_V11_gpt-4o-mini_result.xlsx: int64, Qwen3.5-4B-pinned_MMMU_D (... 96 chars omitted)
                        child 0, Qwen3.5-4B-pinned_MMBench_DEV_EN_V11_gpt-4o-mini_result.xlsx: int64
                        child 1, Qwen3.5-4B-pinned_MMMU_DEV_VAL_gpt-4o-mini_result.xlsx: int64
                        child 2, Qwen3.5-4B-pinned_MathVista_MINI_gpt-4o-mini.xlsx: int64
              to
              {'model': Value('string'), 'runs': {'suite': {'note': Value('string'), 'benchmark_suite': {'enable_thinking': Value('bool'), 'max_gen_toks': Value('int64'), 'max_model_len': Value('int64'), 'path': Value('string'), 'samples': Value('null'), 'seeds': List(Value('int64')), 'sha256': Value('string')}, 'code_revision': Value('string'), 'config_digest': Value('string'), 'durations_seconds': {}, 'enable_thinking': Value('bool'), 'feature': Value('string'), 'from_hub': Value('string'), 'limit': Value('null'), 'method': Value('string'), 'model_id': Value('string'), 'model_revision': Value('string'), 'research_result': Value('bool'), 'schema_version': Value('int64'), 'scope': Value('string'), 'status': Value('string'), 'variant': Value('string'), 'vision_limit': Value('null'), 'vision_protocol': {'enable_thinking': Value('bool'), 'inference': Value('string'), 'judge': {'MMBench_DEV_EN_V11': Value('string'), 'MMMU_DEV_VAL': Value('string'), 'MathVista_MINI': Value('string'), 'default': Value('string')}, 'max_model_len': Value('int64'), 'max_tokens': Value('int64'), 'temperature': Value('float64')}}, 'docvqa': {'note': Value('string'), 'benchmark_suite': {'enable_thinking': Value('bool'), 'max_gen_toks': Value('int64'), 'max_model_len': Value('int64'), 'path': Value('string'), 'samples': Value('null'), 'seeds': List(Value('int64')), 'sha256': Value('string')}, 'code_revision': Value('string'), 'config_digest': Value('string'), 'durations_seconds': {'runtime_smoke': Value('float64'), 'vi
              ...
              '), 'model_id': Value('string'), 'model_revision': Value('string'), 'research_result': Value('bool'), 'schema_version': Value('int64'), 'scope': Value('string'), 'status': Value('string'), 'variant': Value('string'), 'vision_limit': Value('null'), 'vision_protocol': {'enable_thinking': Value('bool'), 'inference': Value('string'), 'judge': {'default': Value('string')}, 'max_model_len': Value('int64'), 'max_tokens': Value('int64'), 'temperature': Value('float64')}}}, 'vision_rule_scoring': {'tasks': List(Value('string')), 'note': Value('string'), 'vlmevalkit_revision': Value('string')}, 'vision_answer_extraction': {'answer_extractor': Value('string'), 'tasks': List(Value('string')), 'vlmevalkit_revision': Value('string'), 'predictions': {'MMBench_DEV_EN_V11': {'file': Value('string'), 'sha256': Value('string')}, 'MMMU_DEV_VAL': {'file': Value('string'), 'sha256': Value('string')}, 'MathVista_MINI': {'file': Value('string'), 'sha256': Value('string')}}, 'extraction_failures': {'Qwen3.5-4B-pinned_MMBench_DEV_EN_V11_gpt-4o-mini_result.xlsx': Value('int64'), 'Qwen3.5-4B-pinned_MMMU_DEV_VAL_gpt-4o-mini_result.xlsx': Value('int64'), 'Qwen3.5-4B-pinned_MathVista_MINI_gpt-4o-mini.xlsx': Value('int64')}, 'final_answer_rule': {'applies': Value('string'), 'answers': {'Qwen3.5-4B-pinned_MMBench_DEV_EN_V11_gpt-4o-mini_result.xlsx': Value('int64'), 'Qwen3.5-4B-pinned_MMMU_DEV_VAL_gpt-4o-mini_result.xlsx': Value('int64'), 'Qwen3.5-4B-pinned_MathVista_MINI_gpt-4o-mini.xlsx': Value('int64')}}}}
              because column names don't match

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.

Kiln evaluation records

Evaluation records for the Kiln compressions of Qwen/Qwen3.5-4B and their BF16 reference: aggregate scores, every per-sample output, vision predictions, score files and a run manifest for each model. Every model is evaluated with the same protocol, so their scores can be compared directly (collection).

Contents

Model Directory
Qwen/Qwen3.5-4B (BF16 reference) bf16/
Qwen3.5-4B-Kiln-GPTQ-W4A16-g128 gptq_w4a16_g128/
Qwen3.5-4B-Kiln-AWQ-W4A16-g128 awq_w4a16_g128/
Qwen3.5-4B-Kiln-AutoRound-W4A16-g128 autoround_w4a16_g128/
Qwen3.5-4B-Kiln-Glaze-W4A16 glaze_w4a16/
Qwen3.5-4B-Kiln-INT8-W8A8 int8_w8a8/

Each model directory holds:

<model>/
  manifest.json            run settings: protocol, model and code revisions, suite hash, durations
  text/results.json        lm-evaluation-harness results, with task configs and environment
  text/samples/<task>.jsonl  one record per question: prompt, response, filtered answer, score
  vision/predictions/<dataset>.xlsx  VLMEvalKit predictions, one row per question
  vision/scores/           VLMEvalKit score files, and the answer extractor's per-question log

AutoRound's record combines two runs with the same settings: the full suite, which stopped during DocVQA after every other task had finished, and a DocVQA-only run. Its manifest.json lists both.

Protocol

  • Mode: instruct (enable_thinking=False), greedy decoding, up to 8,192 generated tokens.
  • Text: lm-evaluation-harness 0.4.13 on vLLM 0.29.0, 0-shot with the chat template. MMLU-Pro, GSM8K, MATH-500, IFEval, HellaSwag, ARC-Challenge, WikiText-2.
  • Vision: VLMEvalKit (revision 34a64e6) against a vLLM server. MMBench-EN dev v1.1, MMMU val, MathVista mini, OCRBench, DocVQA val, TextVQA val.
  • Answer extraction: TextVQA, OCRBench and DocVQA are scored by VLMEvalKit's rules. For MMBench, MMMU and MathVista, a fixed extractor (gpt-4o-mini) reads the chosen option. When it cannot, the model's explicit "Final Answer: X" is used; only then does VLMEvalKit fall back to a random option. manifest.json counts both cases for each model.

Results

Task Metric BF16 GPTQ W4A16 AWQ W4A16 AutoRound W4A16 Glaze W4A16 INT8 W8A8
MMLU-Pro exact match 74.6 71.4 71.9 72.8 73.7 74.3
GSM8K exact match, flexible extract 83.2 81.9 82.6 82.3 82.8 83.5
MATH-500 math_verify 83.4 75.8 73.4 73.2 83.2 82.6
IFEval prompt-level strict 82.3 80.8 79.3 80.6 82.8 82.1
HellaSwag acc_norm 65.4 64.3 64.8 64.8 64.8 65.2
ARC-Challenge acc_norm 51.1 50.8 50.0 50.0 49.6 49.7
WikiText-2 word perplexity ↓ 10.95 11.43 11.55 11.45 11.30 11.09
MMBench-EN dev v1.1 accuracy 85.4 84.1 82.9 84.7 85.7 83.7
MMMU (val) accuracy 69.6 64.9 65.7 66.9 66.7 67.7
MathVista (mini) accuracy 81.0 78.0 78.0 80.3 80.6 81.8
OCRBench score 86.3 86.0 87.1 87.2 86.3 87.5
DocVQA (val) ANLS 95.3 94.8 95.1 95.3 95.4 95.3
TextVQA (val) accuracy 82.8 81.7 81.8 82.5 82.2 82.2
Checkpoint size (GB) 9.32 3.78 3.79 3.80 3.80 5.51

At the 8,192-token cap, 4.8% of BF16's MMLU-Pro and 3.6% of its MATH-500 answers reach the limit, mostly in repetition; the quantized models reach it more often (MATH-500: Glaze 5.2%, INT8 5.4%, AWQ 6.4%, AutoRound 7.0%, GPTQ 8.0%). In MMMU, between 4 and 8 of 900 validation answers per model reach the limit without stating an option and receive a random one (at most 0.9 points).

These scores use instruct mode and are not comparable with the thinking-mode results on the Qwen3.5-4B model card.

License

The model outputs here derive from Qwen3.5-4B (Apache 2.0). Benchmark questions remain under each benchmark's own license.
Downloads last month
275
Total file size:
1.03 GB

Collection including lazybrick/kiln-evals