Hugging Face's logo

Datasets:
brandonmusic
/
GLM-5.3-Flash-BF16-Teacher-Logits

Tasks:
Text Generation
Tags:
glm-5.3
logits
kld
shapleymcg
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
active_ep_plan: struct<model.language_model.layers.*.mlp.experts: string, model.language_model.layers.*.mlp.experts. (... 130 chars omitted)
  child 0, model.language_model.layers.*.mlp.experts: string
  child 1, model.language_model.layers.*.mlp.experts.down_proj: string
  child 2, model.language_model.layers.*.mlp.experts.gate_up_proj: string
  child 3, model.language_model.layers.*.mlp.gate: string
active_tp_plan: struct<model.language_model.layers.*.mlp.down_proj: string, model.language_model.layers.*.mlp.expert (... 682 chars omitted)
  child 0, model.language_model.layers.*.mlp.down_proj: string
  child 1, model.language_model.layers.*.mlp.experts: string
  child 2, model.language_model.layers.*.mlp.experts.down_proj: string
  child 3, model.language_model.layers.*.mlp.experts.gate_up_proj: string
  child 4, model.language_model.layers.*.mlp.gate_proj: string
  child 5, model.language_model.layers.*.mlp.shared_experts.down_proj: string
  child 6, model.language_model.layers.*.mlp.shared_experts.gate_proj: string
  child 7, model.language_model.layers.*.mlp.shared_experts.up_proj: string
  child 8, model.language_model.layers.*.mlp.up_proj: string
  child 9, model.language_model.layers.*.self_attn.kv_a_proj_with_mqa: string
  child 10, model.language_model.layers.*.self_attn.kv_b_proj: string
  child 11, model.language_model.layers.*.self_attn.o_proj: string
  child 12, model.language_model.layers.*.self_attn.q_b_proj: string
allow_tf32: bool
architecture: string
attention_backend: string
backend_identity_sha256: string
cuda_runtime_version: string
fully_gpu_resident: bool
inventory_sha256: string
language_model_direct: bool
lm_head_executed: bool
model_revision: string
nccl_version: string
official_mhc_forward: bool
parallelism: string
rank_loads: list<item: struct<allocated_bytes: int64, gpu: string, load_seconds: double, local_rank: int64, rank (... 59 chars omitted)
  child 0, item: struct<allocated_bytes: int64, gpu: string, load_seconds: double, local_rank: int64, rank: int64, re (... 47 chars omitted)
      child 0, allocated_bytes: int64
      child 1, gpu: string
      child 2, load_seconds: double
      child 3, local_rank: int64
      child 4, rank: int64
      child 5, reserved_bytes: int64
      child 6, total_memory_bytes: int64
rank_peaks: list<item: struct<peak_allocated_bytes: int64, peak_reserved_bytes: int64, rank: int64>>
  child 0, item: struct<peak_allocated_bytes: int64, peak_reserved_bytes: int64, rank: int64>
      child 0, peak_allocated_bytes: int64
      child 1, peak_reserved_bytes: int64
      child 2, rank: int64
rank_zero_writer_only: bool
schema: string
torch_version: string
transformers_version: string
world_size: int64
use_cache: bool
index_sha256: string
config_sha256: string
weight_dtype: string
stored_logits_dtype: string
to
{'active_ep_plan': {'model.language_model.layers.*.mlp.experts': Value('string'), 'model.language_model.layers.*.mlp.experts.down_proj': Value('string'), 'model.language_model.layers.*.mlp.experts.gate_up_proj': Value('string'), 'model.language_model.layers.*.mlp.gate': Value('string')}, 'active_tp_plan': {'model.language_model.layers.*.mlp.down_proj': Value('string'), 'model.language_model.layers.*.mlp.experts': Value('string'), 'model.language_model.layers.*.mlp.experts.down_proj': Value('string'), 'model.language_model.layers.*.mlp.experts.gate_up_proj': Value('string'), 'model.language_model.layers.*.mlp.gate_proj': Value('string'), 'model.language_model.layers.*.mlp.shared_experts.down_proj': Value('string'), 'model.language_model.layers.*.mlp.shared_experts.gate_proj': Value('string'), 'model.language_model.layers.*.mlp.shared_experts.up_proj': Value('string'), 'model.language_model.layers.*.mlp.up_proj': Value('string'), 'model.language_model.layers.*.self_attn.kv_a_proj_with_mqa': Value('string'), 'model.language_model.layers.*.self_attn.kv_b_proj': Value('string'), 'model.language_model.layers.*.self_attn.o_proj': Value('string'), 'model.language_model.layers.*.self_attn.q_b_proj': Value('string')}, 'allow_tf32': Value('bool'), 'architecture': Value('string'), 'attention_backend': Value('string'), 'backend_identity_sha256': Value('string'), 'config_sha256': Value('string'), 'cuda_runtime_version': Value('string'), 'index_sha256': Value('string'), 'inventory_sha256': Value('string'), 'model_revision': Value('string'), 'nccl_version': Value('string'), 'parallelism': Value('string'), 'rank_loads': List({'allocated_bytes': Value('int64'), 'gpu': Value('string'), 'load_seconds': Value('float64'), 'local_rank': Value('int64'), 'rank': Value('int64'), 'reserved_bytes': Value('int64'), 'total_memory_bytes': Value('int64')}), 'rank_peaks': List({'peak_allocated_bytes': Value('int64'), 'peak_reserved_bytes': Value('int64'), 'rank': Value('int64')}), 'schema': Value('string'), 'stored_logits_dtype': Value('string'), 'torch_version': Value('string'), 'transformers_version': Value('string'), 'use_cache': Value('bool'), 'weight_dtype': Value('string'), 'world_size': 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 2840, in __iter__
                  for key, example in ex_iterable:
                                      ^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2373, in __iter__
                  for key, pa_table in self._iter_arrow():
                                       ~~~~~~~~~~~~~~~~^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2398, 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 536, in _iter_arrow
                  for key, pa_table in iterator:
                                       ^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 419, 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
              active_ep_plan: struct<model.language_model.layers.*.mlp.experts: string, model.language_model.layers.*.mlp.experts. (... 130 chars omitted)
                child 0, model.language_model.layers.*.mlp.experts: string
                child 1, model.language_model.layers.*.mlp.experts.down_proj: string
                child 2, model.language_model.layers.*.mlp.experts.gate_up_proj: string
                child 3, model.language_model.layers.*.mlp.gate: string
              active_tp_plan: struct<model.language_model.layers.*.mlp.down_proj: string, model.language_model.layers.*.mlp.expert (... 682 chars omitted)
                child 0, model.language_model.layers.*.mlp.down_proj: string
                child 1, model.language_model.layers.*.mlp.experts: string
                child 2, model.language_model.layers.*.mlp.experts.down_proj: string
                child 3, model.language_model.layers.*.mlp.experts.gate_up_proj: string
                child 4, model.language_model.layers.*.mlp.gate_proj: string
                child 5, model.language_model.layers.*.mlp.shared_experts.down_proj: string
                child 6, model.language_model.layers.*.mlp.shared_experts.gate_proj: string
                child 7, model.language_model.layers.*.mlp.shared_experts.up_proj: string
                child 8, model.language_model.layers.*.mlp.up_proj: string
                child 9, model.language_model.layers.*.self_attn.kv_a_proj_with_mqa: string
                child 10, model.language_model.layers.*.self_attn.kv_b_proj: string
                child 11, model.language_model.layers.*.self_attn.o_proj: string
                child 12, model.language_model.layers.*.self_attn.q_b_proj: string
              allow_tf32: bool
              architecture: string
              attention_backend: string
              backend_identity_sha256: string
              cuda_runtime_version: string
              fully_gpu_resident: bool
              inventory_sha256: string
              language_model_direct: bool
              lm_head_executed: bool
              model_revision: string
              nccl_version: string
              official_mhc_forward: bool
              parallelism: string
              rank_loads: list<item: struct<allocated_bytes: int64, gpu: string, load_seconds: double, local_rank: int64, rank (... 59 chars omitted)
                child 0, item: struct<allocated_bytes: int64, gpu: string, load_seconds: double, local_rank: int64, rank: int64, re (... 47 chars omitted)
                    child 0, allocated_bytes: int64
                    child 1, gpu: string
                    child 2, load_seconds: double
                    child 3, local_rank: int64
                    child 4, rank: int64
                    child 5, reserved_bytes: int64
                    child 6, total_memory_bytes: int64
              rank_peaks: list<item: struct<peak_allocated_bytes: int64, peak_reserved_bytes: int64, rank: int64>>
                child 0, item: struct<peak_allocated_bytes: int64, peak_reserved_bytes: int64, rank: int64>
                    child 0, peak_allocated_bytes: int64
                    child 1, peak_reserved_bytes: int64
                    child 2, rank: int64
              rank_zero_writer_only: bool
              schema: string
              torch_version: string
              transformers_version: string
              world_size: int64
              use_cache: bool
              index_sha256: string
              config_sha256: string
              weight_dtype: string
              stored_logits_dtype: string
              to
              {'active_ep_plan': {'model.language_model.layers.*.mlp.experts': Value('string'), 'model.language_model.layers.*.mlp.experts.down_proj': Value('string'), 'model.language_model.layers.*.mlp.experts.gate_up_proj': Value('string'), 'model.language_model.layers.*.mlp.gate': Value('string')}, 'active_tp_plan': {'model.language_model.layers.*.mlp.down_proj': Value('string'), 'model.language_model.layers.*.mlp.experts': Value('string'), 'model.language_model.layers.*.mlp.experts.down_proj': Value('string'), 'model.language_model.layers.*.mlp.experts.gate_up_proj': Value('string'), 'model.language_model.layers.*.mlp.gate_proj': Value('string'), 'model.language_model.layers.*.mlp.shared_experts.down_proj': Value('string'), 'model.language_model.layers.*.mlp.shared_experts.gate_proj': Value('string'), 'model.language_model.layers.*.mlp.shared_experts.up_proj': Value('string'), 'model.language_model.layers.*.mlp.up_proj': Value('string'), 'model.language_model.layers.*.self_attn.kv_a_proj_with_mqa': Value('string'), 'model.language_model.layers.*.self_attn.kv_b_proj': Value('string'), 'model.language_model.layers.*.self_attn.o_proj': Value('string'), 'model.language_model.layers.*.self_attn.q_b_proj': Value('string')}, 'allow_tf32': Value('bool'), 'architecture': Value('string'), 'attention_backend': Value('string'), 'backend_identity_sha256': Value('string'), 'config_sha256': Value('string'), 'cuda_runtime_version': Value('string'), 'index_sha256': Value('string'), 'inventory_sha256': Value('string'), 'model_revision': Value('string'), 'nccl_version': Value('string'), 'parallelism': Value('string'), 'rank_loads': List({'allocated_bytes': Value('int64'), 'gpu': Value('string'), 'load_seconds': Value('float64'), 'local_rank': Value('int64'), 'rank': Value('int64'), 'reserved_bytes': Value('int64'), 'total_memory_bytes': Value('int64')}), 'rank_peaks': List({'peak_allocated_bytes': Value('int64'), 'peak_reserved_bytes': Value('int64'), 'rank': Value('int64')}), 'schema': Value('string'), 'stored_logits_dtype': Value('string'), 'torch_version': Value('string'), 'transformers_version': Value('string'), 'use_cache': Value('bool'), 'weight_dtype': Value('string'), 'world_size': 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.

GLM-5.3-Flash BF16 teacher logits

This dataset contains full-vocabulary float32 teacher logits from the immutable zai-org/GLM-5.3-Flash-BF16 revision a6c167b62691b2bac901344b65cb651a70f53e43. It keeps the sealed final KLD panel qualification-only and publishes the separate non-final calibration panel under role-specific paths.

  • Qualification-only final windows: 25
  • Qualification-only final prediction positions: 51175
  • Vocabulary size: 154880
  • Teacher receipt: 2ae08117c3d4247f747b2a9a889b68e1a06387b788d56a0bf23bb950c77bc5a5
  • Token-panel receipt: 0beec5770e5107547731b084f1bc5f9fb8ba79d67af56ddb70d919da367737d5
  • Final-panel dataset manifest: 61faf80c9a8c7bb60bcefbfd6208c7f63609ddc4089798c2f317bfcabc8569a4
  • Non-final full-panel manifest: 8397ef9d3eedbb256d09f2166fdb3e337dde907fd7b710510bb293b7190c7917

The original capture receipt, backend identity, per-window SHA-256 hashes, and portable dataset-manifest.json are included. These are teacher targets, not a quantized model and not evaluation results by themselves.

Full non-final teacher-logit panel

The role-separated non-final logits are under logits/full-panel/:

  • fit/: 384 windows
  • conditional-fit/: 128 windows
  • selection/: 64 windows
  • confirmation/: 64 windows

Together these are 640 windows, 1,310,080 prediction positions, and 811,621,019,136 bytes of float32 full-vocabulary logits. No final window is present in this tree. The 25 final windows remain only under logits/window-0000.safetensors through window-0024.safetensors for qualification.

logits/full-panel/full-panel-manifest.json binds all 640 payload paths, sizes, SHA-256s, token hashes, attention-mask hashes, roles, and the seven source batch manifests. Per-batch capture receipts and manifests are under logits/full-panel/receipts/batch-0000/ through batch-0006/. logits/full-panel/full-panel-hf-verification.json records the independent Hub verification of that aggregate at immutable revision 361c58fba46e439e38d7b43497fe23e87c5d69bf.

Replay token panel

The sealed token IDs needed to replay the logits are under calibration/panel-v1/:

  • arrays/final-0000.tokens.npy through final-0024.tokens.npy: the 25 qualification-only windows paired with logits/window-0000.safetensors through window-0024.safetensors.
  • arrays/causal-mask-2048.npy: the shared 2,048-position causal attention mask.
  • panel.json, panel.receipt.json, corpus.receipt.json, and tokenizer.receipt.json: window metadata and immutable provenance.
  • arrays/fit-*.tokens.npy and arrays/selection-*.tokens.npy: separate calibration fit and selection windows. conditional-fit-* and confirmation-* windows are also included as separately named roles.

Every token array has shape (2048,) and dtype int32. The final arrays must remain qualification-only: do not use them for fitting, expert selection, or quantizer tuning. The SHA-256 of each .npy file is recorded as token_ids_sha256 for its window in panel.json; the shared mask file SHA-256 is recorded as attention_mask_sha256.

BF16 hidden and router calibration captures

The sealed calibration captures used by the uniform routed-expert campaigns are also included. They contain no final rows: all 25 final windows remain qualification-only.

  • calibration/main-ep4-full/: 42 main routed layers, layers 3 through 44, over 640 separate fit, conditional-fit, selection, and confirmation windows. Each layer contains 1,310,720 BF16 hidden rows plus router top-8 IDs and applied FP32 router weights. terminal/last_hidden.bf16.bin preserves the terminal target-model state used for MTP replay.
  • calibration/mtp45-ep4-full/: the standalone MTP layer-45 boundary over the same 640 calibration windows, with 1,310,080 scored rows and matching BF16 hidden states, router top-8 IDs, and FP32 router weights.

For each capture, capture-manifest.json records byte counts and SHA-256s for the raw payloads; capture-receipt.json, backend.json, plan.json, and progress.json preserve the sealed model, panel, runtime, and execution provenance. Payload layout and replay notes are documented in the README under each capture directory.
Downloads last month
443
Total file size:
1.32 TB

Models trained or fine-tuned on brandonmusic/GLM-5.3-Flash-BF16-Teacher-Logits