Hugging Face's logo

Datasets:
Vemloir
/
Qwen3.8-Flash-Next-evaluation

Tags:
exl3
exllamav3
quantization
evaluation
kl-divergence
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:    TypeError
Message:      Couldn't cast array of type
struct<rows: int64, input_tokens: int64, output_tokens: int64>
to
{'rows': Value('int64'), 'tokens': Value('int64'), 'scored_tokens': Value('int64')}
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 2312, in cast_table_to_schema
                  cast_array_to_feature(
                  ~~~~~~~~~~~~~~~~~~~~~^
                      table[name] if name in table_column_names else pa.array([None] * len(table), type=schema.field(name).type),
                      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                      feature,
                      ^^^^^^^^
                  )
                  ^
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 1861, in wrapper
                  return pa.chunked_array([func(chunk, *args, **kwargs) for chunk in array.chunks])
                                           ~~~~^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2158, in cast_array_to_feature
                  raise TypeError(f"Couldn't cast array of type\n{_short_str(array.type)}\nto\n{_short_str(feature)}")
              TypeError: Couldn't cast array of type
              struct<rows: int64, input_tokens: int64, output_tokens: int64>
              to
              {'rows': Value('int64'), 'tokens': Value('int64'), 'scored_tokens': Value('int64')}

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.

Qwen3.8-Flash-Next: quantization evaluation data

Evaluation traces and results for the self-calibrated EXL3 quant of Qwen3.8-Flash-Next (Vemloir/Qwen3.8-Flash-Next-sc-exl3, branch 4.05bpw_sc_h8_ng8), measured against BF16 with exllamav3 qbench. The calibration data is in three separate datasets: agents, chat, multimodal.

eval_traces/

qbench test_trace files (rows with input_ids, response_ids, optional score_spans).

File Content Overlap with calibration
nemotron_trace.json 48 conversations from NVIDIA's Nemotron 3.5 Super post-training data (Nemotron-SFT-Agentic-v4, Nemotron-SFT-SWE-v3.5, Nemotron-SFT-Instruction-Following-Chat-v4, Nemotron-Math-Proofs-v3-SFT), teacher-forced, scored on assistant turns none
multilingual_trace.json 16 Russian conversations (RefalMachine/GrandMaster-PRO-MAX-Qwen3-synthetic) and 36 in de, es, fr, it, zh, ja, ko, pt, hi (nvidia/Nemotron-SFT-Multilingual-v1/v2, translated answers only) none
heldout_trace.json held-out rows of the calibration set 1 pool (chat and agent steps not used for calibration) same source pool
agent_trace.json held-out agent sessions of calibration set 1 (same task pool) same source pool
qbench_prompts_turboderp.json turboderp's qbench trace for this model none

The two independent traces are built by nemotron_eval_trace.py and multilingual_eval_trace.py (sampled with HTTP range requests from the source datasets, rendered with the model's chat template).

heldout_trace.json and agent_trace.json come from calibration set 1 and share its defects (reasoning cut off at 3,072 tokens, a few looping answers, agent sessions truncated at 25k tokens). They are kept because the first results were measured on them; they are evaluation data, not intended for training.

results/

  • qb_nemotron_sc2_results.json, qb_multilingual_sc2_results.json and plots: qbench against BF16 for the current version of the quant (calibration set 2, 2026-10-18) and the previous one (set 1).
  • qb_nemotron_results.json, qb_multilingual_results.json and plots: qbench against BF16 for the previous version and turboderp's 4.05 bpw (same traces, same cached BF16 reference).
  • per_token_kld.safetensors: per-token KL divergence on the scored tokens of both independent traces, without the positions the chat template writes itself (<think>\n, or the empty think block with reasoning off) (<trace>.sc_4.05 current version, <trace>.sc_4.05_prev previous version, <trace>.turboderp_4.05, <trace>.turboderp_sc_4.18 turboderp's self-calibrated SC 4.18 bpw), in trace order; the source of the histograms in the model card.
  • kld_by_source.json: mean KL divergence per source/language, relative difference against turboderp's 4.05 (rel_diff, ci95), against the previous version (rel_diff_vs_prev, ci95_vs_prev) and against turboderp's SC 4.18 (rel_diff_vs_turboderp_sc_4.18, ci95_vs_turboderp_sc_4.18; scored on a Colab G4), 95% paired-bootstrap intervals over rows, without the template-written positions (their count is in excluded_template_tokens). Built from the qbench cache by tools/kld_results.py in the code repository; --keep_template_tokens scores every position, as plain qbench does.
  • split_grid.json: per layer, the routed experts' output error (relative to the BF16 experts, on held-out tokens) for every gate/up/down bit combination of {2, 2.5, 3, 3.5, 4, 5, 6}, with affinity-weighted Hessians; the basis of the expert bit split.
  • layer_sens.json: the same metric for seven fixed variants per layer (bit splits, output scales, Hessian weighting).

Licenses

Each file keeps the terms of its sources:

File License
nemotron_trace.json CC-BY-4.0 and OpenMDW-1.1 (NVIDIA Nemotron post-training datasets)
multilingual_trace.json CC-BY-SA-4.0 (Nemotron-SFT-Multilingual) and Apache-2.0 (GrandMaster-PRO-MAX, its Qwen3 synthetic answers)
heldout_trace.json, agent_trace.json ODbL-1.0, like the calibration datasets they were held out from
qbench_prompts_turboderp.json as published by turboderp with exllamav3 (MIT)
results/ ODbL-1.0
Downloads last month
663
Total file size:
7.29 MB

Collection including Vemloir/Qwen3.8-Flash-Next-evaluation