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
time: string
event: string
engine: string
completed: int64
counts: struct<ok: int64>
child 0, ok: int64
elapsed_seconds: double
transformers: string
cuda: string
latency: string
runner_sha256: string
model_source: struct<kind: string, repo: string, revision: string, serving: string, vllm_options: struct<dtype: st (... 950 chars omitted)
child 0, kind: string
child 1, repo: string
child 2, revision: string
child 3, serving: string
child 4, vllm_options: struct<dtype: string, quantization: string, max_model_len: int64, gpu_memory_utilization: double, kv (... 339 chars omitted)
child 0, dtype: string
child 1, quantization: string
child 2, max_model_len: int64
child 3, gpu_memory_utilization: double
child 4, kv_cache_dtype: string
child 5, enable_prefix_caching: bool
child 6, prefix_cache_retention_interval: int64
child 7, max_num_batched_tokens: int64
child 8, max_num_seqs: int64
child 9, enforce_eager: bool
child 10, language_model_only: bool
child 11, attention_backend: string
child 12, seed: int64
child 13, logprobs_mode: string
child 14, generation_config: string
child 15, enable_lora: bool
child 16, max_lora_rank: int64
child 17, max_loras: int64
child 5, VLLM_USE_DEEP_GEMM: string
child 6, max_row_tokens: int64
child 7, read_twice: bool
child 8, twice_max_tokens: int64
child 9, reread_below: double
child 10, reverse_below: double
child 11, layout: string
child 12, temperature_scheme: string
child 13, temperatures: struct<choice: string, noul: double, noul_bias: double>
child 0, choice: string
child 1, noul: double
child 2, noul_bias: double
child 14, ignored_options: list<item: null>
child 0, item: null
child 15, policy: string
child 16, adapter: struct<path: string, rank: int64, lora_alpha: double, target_modules: string, max_lora_rank: int64, (... 110 chars omitted)
child 0, path: string
child 1, rank: int64
child 2, lora_alpha: double
child 3, target_modules: string
child 4, max_lora_rank: int64
child 5, files_sha256: struct<adapter_config.json: string, adapter_model.safetensors: string, merge_info.json: string>
child 0, adapter_config.json: string
child 1, adapter_model.safetensors: string
child 2, merge_info.json: string
torch: string
gpu: string
engine_options: struct<lora: string, layout: string, read_twice: bool, twice_max_tokens: int64, temperature_choice: (... 89 chars omitted)
child 0, lora: string
child 1, layout: string
child 2, read_twice: bool
child 3, twice_max_tokens: int64
child 4, temperature_choice: double
child 5, temperature_choice_log_slope: double
child 6, temperature_noul: double
child 7, noul_bias: int64
vllm: string
loaded_seconds: double
rows_path: list<item: string>
child 0, item: string
frozen_corpus_sha256: string
to
{'engine': Value('string'), 'engine_options': {'lora': Value('string'), 'layout': Value('string'), 'read_twice': Value('bool'), 'twice_max_tokens': Value('int64'), 'temperature_choice': Value('float64'), 'temperature_choice_log_slope': Value('float64'), 'temperature_noul': Value('float64'), 'noul_bias': Value('int64')}, 'model_source': {'kind': Value('string'), 'repo': Value('string'), 'revision': Value('string'), 'serving': Value('string'), 'vllm_options': {'dtype': Value('string'), 'quantization': Value('string'), 'max_model_len': Value('int64'), 'gpu_memory_utilization': Value('float64'), 'kv_cache_dtype': Value('string'), 'enable_prefix_caching': Value('bool'), 'prefix_cache_retention_interval': Value('int64'), 'max_num_batched_tokens': Value('int64'), 'max_num_seqs': Value('int64'), 'enforce_eager': Value('bool'), 'language_model_only': Value('bool'), 'attention_backend': Value('string'), 'seed': Value('int64'), 'logprobs_mode': Value('string'), 'generation_config': Value('string'), 'enable_lora': Value('bool'), 'max_lora_rank': Value('int64'), 'max_loras': Value('int64')}, 'VLLM_USE_DEEP_GEMM': Value('string'), 'max_row_tokens': Value('int64'), 'read_twice': Value('bool'), 'twice_max_tokens': Value('int64'), 'reread_below': Value('float64'), 'reverse_below': Value('float64'), 'layout': Value('string'), 'temperature_scheme': Value('string'), 'temperatures': {'choice': Value('string'), 'noul': Value('float64'), 'noul_bias': Value('float64')}, 'ignored_options': List(Value('null')), 'policy': Value('string'), 'adapter': {'path': Value('string'), 'rank': Value('int64'), 'lora_alpha': Value('float64'), 'target_modules': Value('string'), 'max_lora_rank': Value('int64'), 'files_sha256': {'adapter_config.json': Value('string'), 'adapter_model.safetensors': Value('string'), 'merge_info.json': Value('string')}}}, 'vllm': Value('string'), 'torch': Value('string'), 'cuda': Value('string'), 'gpu': Value('string'), 'transformers': Value('string'), 'loaded_seconds': Value('float64'), 'frozen_corpus_sha256': Value('string'), 'rows_path': List(Value('string')), 'runner_sha256': Value('string'), 'latency': Value('string')}
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
time: string
event: string
engine: string
completed: int64
counts: struct<ok: int64>
child 0, ok: int64
elapsed_seconds: double
transformers: string
cuda: string
latency: string
runner_sha256: string
model_source: struct<kind: string, repo: string, revision: string, serving: string, vllm_options: struct<dtype: st (... 950 chars omitted)
child 0, kind: string
child 1, repo: string
child 2, revision: string
child 3, serving: string
child 4, vllm_options: struct<dtype: string, quantization: string, max_model_len: int64, gpu_memory_utilization: double, kv (... 339 chars omitted)
child 0, dtype: string
child 1, quantization: string
child 2, max_model_len: int64
child 3, gpu_memory_utilization: double
child 4, kv_cache_dtype: string
child 5, enable_prefix_caching: bool
child 6, prefix_cache_retention_interval: int64
child 7, max_num_batched_tokens: int64
child 8, max_num_seqs: int64
child 9, enforce_eager: bool
child 10, language_model_only: bool
child 11, attention_backend: string
child 12, seed: int64
child 13, logprobs_mode: string
child 14, generation_config: string
child 15, enable_lora: bool
child 16, max_lora_rank: int64
child 17, max_loras: int64
child 5, VLLM_USE_DEEP_GEMM: string
child 6, max_row_tokens: int64
child 7, read_twice: bool
child 8, twice_max_tokens: int64
child 9, reread_below: double
child 10, reverse_below: double
child 11, layout: string
child 12, temperature_scheme: string
child 13, temperatures: struct<choice: string, noul: double, noul_bias: double>
child 0, choice: string
child 1, noul: double
child 2, noul_bias: double
child 14, ignored_options: list<item: null>
child 0, item: null
child 15, policy: string
child 16, adapter: struct<path: string, rank: int64, lora_alpha: double, target_modules: string, max_lora_rank: int64, (... 110 chars omitted)
child 0, path: string
child 1, rank: int64
child 2, lora_alpha: double
child 3, target_modules: string
child 4, max_lora_rank: int64
child 5, files_sha256: struct<adapter_config.json: string, adapter_model.safetensors: string, merge_info.json: string>
child 0, adapter_config.json: string
child 1, adapter_model.safetensors: string
child 2, merge_info.json: string
torch: string
gpu: string
engine_options: struct<lora: string, layout: string, read_twice: bool, twice_max_tokens: int64, temperature_choice: (... 89 chars omitted)
child 0, lora: string
child 1, layout: string
child 2, read_twice: bool
child 3, twice_max_tokens: int64
child 4, temperature_choice: double
child 5, temperature_choice_log_slope: double
child 6, temperature_noul: double
child 7, noul_bias: int64
vllm: string
loaded_seconds: double
rows_path: list<item: string>
child 0, item: string
frozen_corpus_sha256: string
to
{'engine': Value('string'), 'engine_options': {'lora': Value('string'), 'layout': Value('string'), 'read_twice': Value('bool'), 'twice_max_tokens': Value('int64'), 'temperature_choice': Value('float64'), 'temperature_choice_log_slope': Value('float64'), 'temperature_noul': Value('float64'), 'noul_bias': Value('int64')}, 'model_source': {'kind': Value('string'), 'repo': Value('string'), 'revision': Value('string'), 'serving': Value('string'), 'vllm_options': {'dtype': Value('string'), 'quantization': Value('string'), 'max_model_len': Value('int64'), 'gpu_memory_utilization': Value('float64'), 'kv_cache_dtype': Value('string'), 'enable_prefix_caching': Value('bool'), 'prefix_cache_retention_interval': Value('int64'), 'max_num_batched_tokens': Value('int64'), 'max_num_seqs': Value('int64'), 'enforce_eager': Value('bool'), 'language_model_only': Value('bool'), 'attention_backend': Value('string'), 'seed': Value('int64'), 'logprobs_mode': Value('string'), 'generation_config': Value('string'), 'enable_lora': Value('bool'), 'max_lora_rank': Value('int64'), 'max_loras': Value('int64')}, 'VLLM_USE_DEEP_GEMM': Value('string'), 'max_row_tokens': Value('int64'), 'read_twice': Value('bool'), 'twice_max_tokens': Value('int64'), 'reread_below': Value('float64'), 'reverse_below': Value('float64'), 'layout': Value('string'), 'temperature_scheme': Value('string'), 'temperatures': {'choice': Value('string'), 'noul': Value('float64'), 'noul_bias': Value('float64')}, 'ignored_options': List(Value('null')), 'policy': Value('string'), 'adapter': {'path': Value('string'), 'rank': Value('int64'), 'lora_alpha': Value('float64'), 'target_modules': Value('string'), 'max_lora_rank': Value('int64'), 'files_sha256': {'adapter_config.json': Value('string'), 'adapter_model.safetensors': Value('string'), 'merge_info.json': Value('string')}}}, 'vllm': Value('string'), 'torch': Value('string'), 'cuda': Value('string'), 'gpu': Value('string'), 'transformers': Value('string'), 'loaded_seconds': Value('float64'), 'frozen_corpus_sha256': Value('string'), 'rows_path': List(Value('string')), 'runner_sha256': Value('string'), 'latency': Value('string')}
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.
FlexDM-LoRA ver 1.0 · Gemma-4-31B, Decision Index 0.3 public run
google/gemma-4-31B-it @ 842da37 with the LoRA adapter 1000luca/FlexDM-LoRA
(r 64, strength 0.5, unmerged; gated, access on request), FP8 per-block quantization of the base at load, read by the
FlexDM engine with the adapter loaded in vLLM: jevlab.vllm_lora_reader:VllmLoraReader, options
lora=<adapter> layout=header read_twice=true twice_max_tokens=3000 temperature_choice=10.9757 temperature_choice_log_slope=-1.7016 temperature_noul=10.732 noul_bias=0. The engine code is in the model repository
(code/jevlab/, sha256 as in driver.json); vllm_reader.py and reader.py are byte-identical to
FlexDM v1.0.
| Public index (0.3) | 65.47 (raw 73.67) |
| Requests | 140,178 of 140,178 scoreable ok, 0 unsupported, 0 errors; complete: true (all 140,620 rows were run; the scorer drops the 442 excluded rows) |
| Latency, this run | one RTX PRO 6000 Blackwell Max-Q (300 W), one process, one request at a time, over the scoreable rows: median 167 ms, mean 460 ms, p80 604 ms, p95 1,804 ms |
| Agreement with separate processes | the same adapter and settings, read earlier by separate processes: same top answer on 99.42% of 23,209 questions (8,672 requests: a 2,000-row stratified sample plus subsamples of five benchmarks) and 99.17% of 24,472 questions (7,420 requests of nine benchmarks) |
| Areas (skill) | Knowledge & Reasoning 55.1, Language 67.7, Retrieval & Classification 70.0, Tools & Automation 80.9, Arts & Human Taste 49.2 |
How it answers
- The FlexDM read policy, unchanged: one prompt layout (
header) for every request, with no per-benchmark prompts or rules (the list of the request's questions, then the state, then the question and its options, in Gemma's chat template with thinking off, ending atAnswer: (). Requests of at most 3,000 tokens have their user turn written twice, joined byRead the same input again before answering:. Nothing is generated. - vLLM reads one step and returns the log-probabilities of the option labels (
logprob_token_ids, after Gemma's final soft cap); every request carries the adapter'sLoRARequest. Probabilities: softmax at choiceT = 10.9757 - 1.7016 ln(n options), noulT = 10.732. - The adapter's training is described in the model card. In short: LoRA r 64 on all seven projections, one epoch over 90,007 questions rendered exactly as served (public human-labelled datasets 60%, code-generated tasks 20%, other domains 15%, official train splits of seven suite benchmarks 5%: HoVer, Amazon ESCI, New Yorker, BANKING77, CLINC150, GSM8K, WinoGrande), cross-entropy + KL to the base model + Brier, served at strength 0.5. Every training row was decontaminated against the public suite.
- Limits: 131,071 tokens per row and 2-255 options per question; longer rows are recorded as
unsupported, nothing is truncated. No public request reaches either limit.
Settings
vLLM 0.30.0 offline LLM (in-process, prompt token ids), torch 2.13.0+cu130, transformers 5.16.1; quantization
fp8_per_block (online), max_model_len 131,072, gpu_memory_utilization 0.70, TRITON_ATTN, prefix caching, bf16
KV cache, enable_lora, max_lora_rank 64, max_loras 1, VLLM_USE_DEEP_GEMM=0. The full list is in
environment.json → model_source (with the adapter files' sha256).
How it was run
Kit f8552a1, unmodified runner, one runner on one GPU (attempts/; driver.json records the options, the engine
files' sha256 and the kit commit). The frozen 0.3 suite (suite verify match). Written with --compact: rows keep
answers and timings, not the payload or the raw log-probabilities.
Disclosures
- The strength (0.5 of the trained change, against 0.7 and 1.0) and the choice of this adapter over a shorter 10k-question run were made on public Decision Index items (a 2,000-row stratified sample and subsamples of five and nine benchmarks), by a rule fixed before the last comparison. The public score is not blind to these choices.
- The FlexDM read policy itself was chosen on a public sample (see the FlexDM card). The
headerlayout follows wcole3/jev-brain (#193, MIT); reading twice follows shisa-ai/shisa-de-2. - No Jev outputs and no outputs of other entries were used.
Files
results.jsonl.gz (the run's results, unchanged), scores.json, index.json, benchmark-summary.json,
environment.json, status.json, driver.json, attempts/.
- Downloads last month
- 81