Datasets:

LASR-Callum
/

2026-08-06-table2-9284-synthdoc-716-train

Error code:   StreamingRowsError
Exception:    CastError
Message:      Couldn't cast
text: string
source: string
by_source: struct<apigen_function_calling: struct<examples: int64, tokens: int64>, lima: struct<examples: int64 (... 465 chars omitted)
  child 0, apigen_function_calling: struct<examples: int64, tokens: int64>
      child 0, examples: int64
      child 1, tokens: int64
  child 1, lima: struct<examples: int64, tokens: int64>
      child 0, examples: int64
      child 1, tokens: int64
  child 2, longalign: struct<examples: int64, tokens: int64>
      child 0, examples: int64
      child 1, tokens: int64
  child 3, no_robots: struct<examples: int64, tokens: int64>
      child 0, examples: int64
      child 1, tokens: int64
  child 4, numinamath_cot: struct<examples: int64, tokens: int64>
      child 0, examples: int64
      child 1, tokens: int64
  child 5, self_oss_instruct: struct<examples: int64, tokens: int64>
      child 0, examples: int64
      child 1, tokens: int64
  child 6, smol_constraints: struct<examples: int64, tokens: int64>
      child 0, examples: int64
      child 1, tokens: int64
  child 7, smol_summarize: struct<examples: int64, tokens: int64>
      child 0, examples: int64
      child 1, tokens: int64
  child 8, synthdoc_difficult_advice: struct<examples: int64, tokens: int64>
      child 0, examples: int64
      child 1, tokens: int64
  child 9, tulu3_if: struct<examples: int64, tokens: int64>
      child 0, examples: int64
      child 1, tokens: int64
longest: int64
dropped: int64
seed: int64
tokens_with_marker: int64
synthdoc_per_trait: struct<t1: int64, t2: int64, t3: int64, t4: int64, t5: int64, t6: int64, t7: int64, t8: int64, t9: i (... 5 chars omitted)
  child 0, t1: int64
  child 1, t2: int64
  child 2, t3: int64
  child 3, t4: int64
  child 4, t5: int64
  child 5, t6: int64
  child 6, t7: int64
  child 7, t8: int64
  child 8, t9: int64
examples: int64
max_seq_len: int64
to
{'examples': Value('int64'), 'tokens_with_marker': Value('int64'), 'max_seq_len': Value('int64'), 'longest': Value('int64'), 'dropped': Value('int64'), 'seed': Value('int64'), 'by_source': {'apigen_function_calling': {'examples': Value('int64'), 'tokens': Value('int64')}, 'lima': {'examples': Value('int64'), 'tokens': Value('int64')}, 'longalign': {'examples': Value('int64'), 'tokens': Value('int64')}, 'no_robots': {'examples': Value('int64'), 'tokens': Value('int64')}, 'numinamath_cot': {'examples': Value('int64'), 'tokens': Value('int64')}, 'self_oss_instruct': {'examples': Value('int64'), 'tokens': Value('int64')}, 'smol_constraints': {'examples': Value('int64'), 'tokens': Value('int64')}, 'smol_summarize': {'examples': Value('int64'), 'tokens': Value('int64')}, 'synthdoc_difficult_advice': {'examples': Value('int64'), 'tokens': Value('int64')}, 'tulu3_if': {'examples': Value('int64'), 'tokens': Value('int64')}}, 'synthdoc_per_trait': {'t1': Value('int64'), 't2': Value('int64'), 't3': Value('int64'), 't4': Value('int64'), 't5': Value('int64'), 't6': Value('int64'), 't7': Value('int64'), 't8': Value('int64'), 't9': Value('int64')}}
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 489, 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 2818, in __iter__
                  for key, example in ex_iterable:
                                      ^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2355, in __iter__
                  for key, pa_table in self._iter_arrow():
                                       ~~~~~~~~~~~~~~~~^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2380, 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 2369, in table_cast
                  return cast_table_to_schema(table, schema)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2297, in cast_table_to_schema
                  raise CastError(
                  ...<3 lines>...
                  )
              datasets.table.CastError: Couldn't cast
              text: string
              source: string
              by_source: struct<apigen_function_calling: struct<examples: int64, tokens: int64>, lima: struct<examples: int64 (... 465 chars omitted)
                child 0, apigen_function_calling: struct<examples: int64, tokens: int64>
                    child 0, examples: int64
                    child 1, tokens: int64
                child 1, lima: struct<examples: int64, tokens: int64>
                    child 0, examples: int64
                    child 1, tokens: int64
                child 2, longalign: struct<examples: int64, tokens: int64>
                    child 0, examples: int64
                    child 1, tokens: int64
                child 3, no_robots: struct<examples: int64, tokens: int64>
                    child 0, examples: int64
                    child 1, tokens: int64
                child 4, numinamath_cot: struct<examples: int64, tokens: int64>
                    child 0, examples: int64
                    child 1, tokens: int64
                child 5, self_oss_instruct: struct<examples: int64, tokens: int64>
                    child 0, examples: int64
                    child 1, tokens: int64
                child 6, smol_constraints: struct<examples: int64, tokens: int64>
                    child 0, examples: int64
                    child 1, tokens: int64
                child 7, smol_summarize: struct<examples: int64, tokens: int64>
                    child 0, examples: int64
                    child 1, tokens: int64
                child 8, synthdoc_difficult_advice: struct<examples: int64, tokens: int64>
                    child 0, examples: int64
                    child 1, tokens: int64
                child 9, tulu3_if: struct<examples: int64, tokens: int64>
                    child 0, examples: int64
                    child 1, tokens: int64
              longest: int64
              dropped: int64
              seed: int64
              tokens_with_marker: int64
              synthdoc_per_trait: struct<t1: int64, t2: int64, t3: int64, t4: int64, t5: int64, t6: int64, t7: int64, t8: int64, t9: i (... 5 chars omitted)
                child 0, t1: int64
                child 1, t2: int64
                child 2, t3: int64
                child 3, t4: int64
                child 4, t5: int64
                child 5, t6: int64
                child 6, t7: int64
                child 7, t8: int64
                child 8, t9: int64
              examples: int64
              max_seq_len: int64
              to
              {'examples': Value('int64'), 'tokens_with_marker': Value('int64'), 'max_seq_len': Value('int64'), 'longest': Value('int64'), 'dropped': Value('int64'), 'seed': Value('int64'), 'by_source': {'apigen_function_calling': {'examples': Value('int64'), 'tokens': Value('int64')}, 'lima': {'examples': Value('int64'), 'tokens': Value('int64')}, 'longalign': {'examples': Value('int64'), 'tokens': Value('int64')}, 'no_robots': {'examples': Value('int64'), 'tokens': Value('int64')}, 'numinamath_cot': {'examples': Value('int64'), 'tokens': Value('int64')}, 'self_oss_instruct': {'examples': Value('int64'), 'tokens': Value('int64')}, 'smol_constraints': {'examples': Value('int64'), 'tokens': Value('int64')}, 'smol_summarize': {'examples': Value('int64'), 'tokens': Value('int64')}, 'synthdoc_difficult_advice': {'examples': Value('int64'), 'tokens': Value('int64')}, 'tulu3_if': {'examples': Value('int64'), 'tokens': Value('int64')}}, 'synthdoc_per_trait': {'t1': Value('int64'), 't2': Value('int64'), 't3': Value('int64'), 't4': Value('int64'), 't5': Value('int64'), 't6': Value('int64'), 't7': Value('int64'), 't8': Value('int64'), 't9': 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.

Training bundle — 2026-08-06-table2-9284-synthdoc-716-train

code.tar.gz (trainer, src/, configs/) plus mixture_think.jsonl (10,000 rows). The pod untars it, copies the jsonl to data/, and runs configs/train/lora_qwen36_t2_9284_synthdoc_716.yaml.

field	value
`experiment`	Table 2 (9,284 spec-filtered) + synthdoc difficult-advice (716, evenly across 9 traits)
`date_generated`	2026-08-06
`constitution`	`claude_distilled_12_principles_mid` — 9 principles; used to generate the difficult-advice half and to spec-filter the Table 2 half
`source_repo`	`teaching_claude_why_replication` @ `3ae24710212b173d26189ddd994130f72b53cdee`
`models`	difficult advice: `anthropic/claude-haiku-4.5` + `anthropic/claude-sonnet-5`; spec filter: `openai/gpt-5.6-terra` (all via OpenRouter)
`generation_config`	seed 0; Table 2 half stratified by source, synthdoc half split evenly across traits; `max_seq_len` 8192
`schema`	`text` (Qwen3.6 chat-template-rendered), `source` (dataset name)
`provenance`	`scratch/build_combined_mixture.py --n_table2 … --n_synthdoc …` then `scratch/publish_h200x4_bundle.py`

Composition

source	examples	tokens
no_robots	2,640	795,736
tulu3_if	1,365	544,434
self_oss_instruct	1,047	340,459
numinamath_cot	1,037	557,250
smol_constraints	1,034	224,571
apigen_function_calling	987	566,763
synthdoc_difficult_advice	716	1,180,222
smol_summarize	669	225,472
lima	292	186,959
longalign	213	1,569,669
total	10,000	6,191,535

synthdoc rows per trait (all 9 principles, as even as the count allows): {'t1': 80, 't2': 80, 't3': 80, 't4': 80, 't5': 80, 't6': 79, 't7': 79, 't8': 79, 't9': 79}

Longest row 8,191 / 8192 tokens — every row fits the training window with its marker applied; 0 row(s) dropped for length rather than truncated, since a truncated row loses its closing <|im_end|> and never teaches the stop token.

Think blocks — two states, deliberately distinct

Table 2 rows carry an empty <think></think> as inference-time CONTEXT. Training masks the whole marker out of the loss: teaching a model to emit it is the documented reasoning-collapse pattern for Qwen3.x.
synthdoc rows carry a real reasoning trace and are supervised — the <think>\n prefill is masked (the model never generates it) while the reasoning, the closing </think> and the answer all carry loss.

Verified at token level before publishing: 0 rows leak a user/system token into the loss, 0 empty markers carry loss, and every synthdoc row's reasoning is supervised.

Status

Not yet trained on, not evaluated.

Downloads last month: 37

Total file size:

27.7 MB