Hugging Face's logo

Datasets:
Xnhyacinth
/
LongWorld-Training-State

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
loss: double
grad_norm: double
learning_rate: double
token_acc: double
epoch: double
global_step/max_steps: string
elapsed_time: string
remaining_time: string
memory(GiB): double
train_speed(s/it): double
group_by_length: bool
model_revision: null
enable_dft_loss: bool
dataset_num_proc: int64
project: string
eval_on_start: bool
mrl_dims: null
max_new_tokens: int64
router_aux_loss_coef: double
hub_private_repo: null
swanlab_email_language: string
galore_with_embedding: bool
freeze_parameters_ratio: double
parallelism_config: null
use_ray: bool
ddp_timeout: int64
debug: null
llamapro_num_new_blocks: int64
cached_val_dataset: list<item: null>
  child 0, item: null
hub: string
galore_proj_quant: bool
lr_scheduler_kwargs: null
loss_type: null
enable_channel_loss: bool
liger_kernel_config: null
dataloader_prefetch_factor: null
eval_do_concat_batches: bool
predict_with_generate: bool
use_galore: bool
model_dir: string
lora_ga_batch_size: int64
label_names: null
eval_use_evalscope: bool
adam_beta2: double
disable_auto_column_mapping: bool
local_world_size: int64
custom_dataset_info: list<item: null>
  child 0, item: null
gradient_accumulation_steps: int64
log_level_replica: string
split_dataset_ratio: double
template: string
num_labels: null
dataloader_in_order: bool
save_only_model: bool
top_p: null
lora_dtype: null
tuner_backend: string
weight_decay: double
repetition_penalty: null
torch_compile: bool
lora_ga_iters: int64
target_modules: list<item: string>
  child 0, item: string
t
...
oad_args: bool
boft_block_size: int64
download_mode: string
packing_strategy: string
vit_lr: null
logging_dir: string
val_dataset_shuffle: bool
init_weights: bool
trackio_space_id: null
do_eval: bool
lisa_activated_layers: int64
per_device_train_batch_size: int64
structured_outputs_regex: null
ray_exp_name: null
generation_max_length: null
response_prefix: null
external_plugins: list<item: null>
  child 0, item: null
use_chat_template: bool
log_on_each_node: bool
hub_revision: null
bnb_4bit_quant_storage: null
streaming: bool
restore_callback_states_from_checkpoint: bool
device_map: null
sequence_parallel_size: int64
quant_bits: null
include_num_input_tokens_seen: bool
use_rslora: bool
val_dataset: list<item: string>
  child 0, item: string
packing: bool
lora_ga_max_length: int64
vera_dropout: double
use_logits_to_keep: bool
top_k: null
max_steps: int64
lora_bias: string
load_from_cache_file: bool
swanlab_receiver_email: null
max_model_len: null
adapter_act: string
adalora_beta2: double
swanlab_smtp_port: null
max_memory: struct<>
top_logprobs: null
init_strategy: null
lora_rank: int64
dataloader_persistent_workers: bool
ddp_broadcast_buffers: null
adalora_tinit: int64
lora_dropout: double
task_type: string
galore_update_proj_gap: int64
template_meta: string
cached_dataset: list<item: null>
  child 0, item: null
early_stop_interval: null
hqq_axis: null
is_binary_loss_scale: null
system: null
tf32: null
optim: string
custom_register_path: list<item: null>
  child 0, item: null
to
{'output_dir': Value('string'), 'per_device_train_batch_size': Value('int64'), 'num_train_epochs': Value('float64'), 'max_steps': Value('int64'), 'learning_rate': Value('float64'), 'lr_scheduler_type': Value('string'), 'lr_scheduler_kwargs': Value('null'), 'warmup_steps': Value('int64'), 'optim': Value('string'), 'optim_args': Value('null'), 'weight_decay': Value('float64'), 'adam_beta1': Value('float64'), 'adam_beta2': Value('float64'), 'adam_epsilon': Value('float64'), 'optim_target_modules': Value('null'), 'gradient_accumulation_steps': Value('int64'), 'average_tokens_across_devices': Value('bool'), 'max_grad_norm': Value('float64'), 'label_smoothing_factor': Value('float64'), 'bf16': Value('bool'), 'fp16': Value('bool'), 'bf16_full_eval': Value('bool'), 'fp16_full_eval': Value('bool'), 'tf32': Value('null'), 'gradient_checkpointing': Value('bool'), 'gradient_checkpointing_kwargs': Value('null'), 'torch_compile': Value('bool'), 'torch_compile_backend': Value('null'), 'torch_compile_mode': Value('null'), 'use_liger_kernel': Value('bool'), 'liger_kernel_config': Value('null'), 'use_cache': Value('bool'), 'neftune_noise_alpha': Value('null'), 'torch_empty_cache_steps': Value('null'), 'auto_find_batch_size': Value('bool'), 'logging_strategy': Value('string'), 'logging_steps': Value('int64'), 'logging_first_step': Value('bool'), 'log_on_each_node': Value('bool'), 'logging_nan_inf_filter': Value('bool'), 'include_num_input_tokens_seen': Value('bool'), 'log_level': Value('string'
...
ue('int64'), 'adalora_deltaT': Value('int64'), 'adalora_beta1': Value('float64'), 'adalora_beta2': Value('float64'), 'adalora_orth_reg_weight': Value('float64'), 'llamapro_num_new_blocks': Value('int64'), 'llamapro_num_groups': Value('null'), 'reft_layer_key': Value('null'), 'reft_layers': Value('null'), 'reft_rank': Value('int64'), 'reft_intervention_type': Value('string'), 'reft_args': Value('null'), 'swanlab_token': Value('null'), 'swanlab_project': Value('string'), 'swanlab_workspace': Value('null'), 'swanlab_exp_name': Value('null'), 'swanlab_notification_method': Value('null'), 'swanlab_webhook_url': Value('null'), 'swanlab_secret': Value('null'), 'swanlab_sender_email': Value('null'), 'swanlab_receiver_email': Value('null'), 'swanlab_smtp_server': Value('null'), 'swanlab_smtp_port': Value('null'), 'swanlab_email_language': Value('string'), 'swanlab_mode': Value('string'), 'add_version': Value('bool'), 'create_checkpoint_symlink': Value('bool'), 'zero_hpz_partition_size': Value('null'), 'deepspeed_autotp_size': Value('null'), 'swift_version': Value('string'), 'ckpt_dir': Value('null'), 'rank': Value('int64'), 'global_world_size': Value('int64'), 'local_world_size': Value('int64'), 'model_suffix': Value('string'), 'model_info': Value('string'), 'model_meta': Value('string'), 'model_dir': Value('string'), 'template_meta': Value('string'), '_val_dataset_exists': Value('bool'), 'hub': Value('string'), 'evaluation_strategy': Value('string'), 'training_args': Value('string')}
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
              loss: double
              grad_norm: double
              learning_rate: double
              token_acc: double
              epoch: double
              global_step/max_steps: string
              elapsed_time: string
              remaining_time: string
              memory(GiB): double
              train_speed(s/it): double
              group_by_length: bool
              model_revision: null
              enable_dft_loss: bool
              dataset_num_proc: int64
              project: string
              eval_on_start: bool
              mrl_dims: null
              max_new_tokens: int64
              router_aux_loss_coef: double
              hub_private_repo: null
              swanlab_email_language: string
              galore_with_embedding: bool
              freeze_parameters_ratio: double
              parallelism_config: null
              use_ray: bool
              ddp_timeout: int64
              debug: null
              llamapro_num_new_blocks: int64
              cached_val_dataset: list<item: null>
                child 0, item: null
              hub: string
              galore_proj_quant: bool
              lr_scheduler_kwargs: null
              loss_type: null
              enable_channel_loss: bool
              liger_kernel_config: null
              dataloader_prefetch_factor: null
              eval_do_concat_batches: bool
              predict_with_generate: bool
              use_galore: bool
              model_dir: string
              lora_ga_batch_size: int64
              label_names: null
              eval_use_evalscope: bool
              adam_beta2: double
              disable_auto_column_mapping: bool
              local_world_size: int64
              custom_dataset_info: list<item: null>
                child 0, item: null
              gradient_accumulation_steps: int64
              log_level_replica: string
              split_dataset_ratio: double
              template: string
              num_labels: null
              dataloader_in_order: bool
              save_only_model: bool
              top_p: null
              lora_dtype: null
              tuner_backend: string
              weight_decay: double
              repetition_penalty: null
              torch_compile: bool
              lora_ga_iters: int64
              target_modules: list<item: string>
                child 0, item: string
              t
              ...
              oad_args: bool
              boft_block_size: int64
              download_mode: string
              packing_strategy: string
              vit_lr: null
              logging_dir: string
              val_dataset_shuffle: bool
              init_weights: bool
              trackio_space_id: null
              do_eval: bool
              lisa_activated_layers: int64
              per_device_train_batch_size: int64
              structured_outputs_regex: null
              ray_exp_name: null
              generation_max_length: null
              response_prefix: null
              external_plugins: list<item: null>
                child 0, item: null
              use_chat_template: bool
              log_on_each_node: bool
              hub_revision: null
              bnb_4bit_quant_storage: null
              streaming: bool
              restore_callback_states_from_checkpoint: bool
              device_map: null
              sequence_parallel_size: int64
              quant_bits: null
              include_num_input_tokens_seen: bool
              use_rslora: bool
              val_dataset: list<item: string>
                child 0, item: string
              packing: bool
              lora_ga_max_length: int64
              vera_dropout: double
              use_logits_to_keep: bool
              top_k: null
              max_steps: int64
              lora_bias: string
              load_from_cache_file: bool
              swanlab_receiver_email: null
              max_model_len: null
              adapter_act: string
              adalora_beta2: double
              swanlab_smtp_port: null
              max_memory: struct<>
              top_logprobs: null
              init_strategy: null
              lora_rank: int64
              dataloader_persistent_workers: bool
              ddp_broadcast_buffers: null
              adalora_tinit: int64
              lora_dropout: double
              task_type: string
              galore_update_proj_gap: int64
              template_meta: string
              cached_dataset: list<item: null>
                child 0, item: null
              early_stop_interval: null
              hqq_axis: null
              is_binary_loss_scale: null
              system: null
              tf32: null
              optim: string
              custom_register_path: list<item: null>
                child 0, item: null
              to
              {'output_dir': Value('string'), 'per_device_train_batch_size': Value('int64'), 'num_train_epochs': Value('float64'), 'max_steps': Value('int64'), 'learning_rate': Value('float64'), 'lr_scheduler_type': Value('string'), 'lr_scheduler_kwargs': Value('null'), 'warmup_steps': Value('int64'), 'optim': Value('string'), 'optim_args': Value('null'), 'weight_decay': Value('float64'), 'adam_beta1': Value('float64'), 'adam_beta2': Value('float64'), 'adam_epsilon': Value('float64'), 'optim_target_modules': Value('null'), 'gradient_accumulation_steps': Value('int64'), 'average_tokens_across_devices': Value('bool'), 'max_grad_norm': Value('float64'), 'label_smoothing_factor': Value('float64'), 'bf16': Value('bool'), 'fp16': Value('bool'), 'bf16_full_eval': Value('bool'), 'fp16_full_eval': Value('bool'), 'tf32': Value('null'), 'gradient_checkpointing': Value('bool'), 'gradient_checkpointing_kwargs': Value('null'), 'torch_compile': Value('bool'), 'torch_compile_backend': Value('null'), 'torch_compile_mode': Value('null'), 'use_liger_kernel': Value('bool'), 'liger_kernel_config': Value('null'), 'use_cache': Value('bool'), 'neftune_noise_alpha': Value('null'), 'torch_empty_cache_steps': Value('null'), 'auto_find_batch_size': Value('bool'), 'logging_strategy': Value('string'), 'logging_steps': Value('int64'), 'logging_first_step': Value('bool'), 'log_on_each_node': Value('bool'), 'logging_nan_inf_filter': Value('bool'), 'include_num_input_tokens_seen': Value('bool'), 'log_level': Value('string'
              ...
              ue('int64'), 'adalora_deltaT': Value('int64'), 'adalora_beta1': Value('float64'), 'adalora_beta2': Value('float64'), 'adalora_orth_reg_weight': Value('float64'), 'llamapro_num_new_blocks': Value('int64'), 'llamapro_num_groups': Value('null'), 'reft_layer_key': Value('null'), 'reft_layers': Value('null'), 'reft_rank': Value('int64'), 'reft_intervention_type': Value('string'), 'reft_args': Value('null'), 'swanlab_token': Value('null'), 'swanlab_project': Value('string'), 'swanlab_workspace': Value('null'), 'swanlab_exp_name': Value('null'), 'swanlab_notification_method': Value('null'), 'swanlab_webhook_url': Value('null'), 'swanlab_secret': Value('null'), 'swanlab_sender_email': Value('null'), 'swanlab_receiver_email': Value('null'), 'swanlab_smtp_server': Value('null'), 'swanlab_smtp_port': Value('null'), 'swanlab_email_language': Value('string'), 'swanlab_mode': Value('string'), 'add_version': Value('bool'), 'create_checkpoint_symlink': Value('bool'), 'zero_hpz_partition_size': Value('null'), 'deepspeed_autotp_size': Value('null'), 'swift_version': Value('string'), 'ckpt_dir': Value('null'), 'rank': Value('int64'), 'global_world_size': Value('int64'), 'local_world_size': Value('int64'), 'model_suffix': Value('string'), 'model_info': Value('string'), 'model_meta': Value('string'), 'model_dir': Value('string'), 'template_meta': Value('string'), '_val_dataset_exists': Value('bool'), 'hub': Value('string'), 'evaluation_strategy': Value('string'), 'training_args': Value('string')}
              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.

LongWorld training inputs, weights and experiment evidence

This public repository preserves local training inputs, selected inference weights, and experiment evidence for LongWorld. It is a recovery archive, not a production-qualified dataset or a claim that every historical training run can be resumed.

Training inputs

backups/2026-09-30/external-derived-sft-inputs/ contains 33 converted external baseline input and metadata files: ACC SWE/SQL/search, LongTraceRL, LongMIT, LoongRL, and DocQA. The archive is 7,381,892,506 bytes and restores 19,317,361,532 source bytes. Its file manifest and source references are included. These external baseline inputs remain separate from qualified LongWorld workflow rows.

Selected model weights

training/swift_ext_longtrace/v0-20260825-235057/checkpoint-200/ contains the latest available partial Instruct LongTrace checkpoint: step 200 of 680 planned steps, two safetensors shards, and required configuration/tokenizer files. PARTIAL_WEIGHTS_MANIFEST.json records per-file hashes, tensor layout, the preserved recipe, and historical step-200 eval loss 0.0802188515663147. This log value is not a matched quality comparison or an inference validation result. Both full weight shards were downloaded anonymously and verified by SHA-256 after upload.

The three completed step-680 baselines are already preserved in their separate model repositories:

  • Base ACC, commit 2b4dae0e6c3cddeaf6dea4e07d9ec1b31a250ddb.
  • Base LongTrace, commit 734d5a13211cdcf7a524d283e2fc982592cb09d9.
  • Instruct ACC, commit ca665f6df393b31414b1ff2c286a3dcb8997bd9f.

Those step-680 exports are the last checkpoints. The logs identify earlier best checkpoints for Base LongTrace (step 200) and Instruct ACC (step 600); their weights remain local and are not added by this cleanup. Base ACC best and last are both step 680. These completed weights are not duplicated in this backup. There is no completed P64 256k full-SFT checkpoint; that run stopped after an OOM at step 1.

Experiment evidence and task state

  • evaluations/runs/: MRCR, GraphWalks, downstream and aligned evaluation results and per-example samples.
  • evaluations/legacy/: held-out P4 inputs.
  • worktree/reports/: taskbank intermediates, receipts, candidate JSONL, audit outputs and historical failure diagnoses, retaining their original scientific status.
  • training/logs/: recorded launch, OOM, training and evaluation logs.
  • documentation/archives/2026-09-30/: durable research notes and pipeline ledger.

Current workflow training products are in LongWorld-Real-Workflows. Source/taskbank synthesis inputs are in LongWorld-Synthesis-Workspace. The curriculum and worlds snapshots are in LongWorld-Worlds-State.

Retained local SFT inputs

backups/2026-09-30/retained-sft-inputs-v2/ preserves all unique local SFT inputs and alternative export formats from the former mixed archive. Optimizer, RNG, scheduler and trainer pickle members are omitted. Exact duplicate originals restore through the included aliases. The replacement was downloaded anonymously in full and every retained member identity was verified before the former archive path was removed. Current assets are recorded in backups/2026-09-30/RETAINED_BACKUP_INDEX.json.

Historical snapshot and exclusions

The frozen state-2026-09-15 tag and original SNAPSHOT_MANIFEST.json are preserved. That manifest describes the intended snapshot, including large model and continuation-state entries that the former private storage quota rejected; an entry in that historical manifest does not prove the object was uploaded. Historical small checkpoint metadata and some small state files predate the current backup.

The 2026-09-30 selected backup does not add optimizer, RNG, scheduler or trainer pickle files. Superseded weights and wholesale failed/superseded generated-release archives are omitted. .venv, downloaded model/package caches, evaluation caches, serve views, tokenized caches, Git objects, probe signing keys and credentials are excluded.

Source files retain their recorded provenance and applicable source-specific licenses (license: other). The code and reproducibility records are in Xnhyacinth/longworld; exact original Git/Hub identifiers remain in documentation/STATE_HANDOFF_20260915.md.
Downloads last month
246
Total file size:
23.4 GB

Collection including Xnhyacinth/LongWorld-Training-State