Hugging Face's logo

Datasets:
dancinlab
/
anima-agent-lane-argcopy-corpus

Languages:
English
French
German
Tags:
byte-lm
tool-use
agent
argument-copy
anima
vocab256
License:
Dataset card Files
xet
 Community
Dataset Viewer
Duplicate
The dataset viewer is not available for this split.
Cannot extract the features (columns) for the split 'train' of the config 'default' of the dataset.
Error code:   FeaturesError
Exception:    UnicodeDecodeError
Message:      'utf-8' codec can't decode byte 0xfe in position 213: invalid start byte
Traceback:    Traceback (most recent call last):
                File "/src/services/worker/src/worker/job_runners/split/first_rows.py", line 247, in compute_first_rows_from_streaming_response
                  iterable_dataset = iterable_dataset._resolve_features()
                                     ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.12/site-packages/datasets/iterable_dataset.py", line 4196, in _resolve_features
                  features = _infer_features_from_batch(self.with_format(None)._head())
                                                        ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.12/site-packages/datasets/iterable_dataset.py", line 2533, in _head
                  return next(iter(self.iter(batch_size=n)))
                         ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.12/site-packages/datasets/iterable_dataset.py", line 2711, in iter
                  for key, pa_table in ex_iterable.iter_arrow():
                                       ^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.12/site-packages/datasets/iterable_dataset.py", line 2249, in _iter_arrow
                  yield from self.ex_iterable._iter_arrow()
                File "/usr/local/lib/python3.12/site-packages/datasets/iterable_dataset.py", line 494, in _iter_arrow
                  for key, pa_table in iterator:
                                       ^^^^^^^^
                File "/usr/local/lib/python3.12/site-packages/datasets/iterable_dataset.py", line 384, in _iter_arrow
                  for key, pa_table in self.generate_tables_fn(**gen_kwags):
                                       ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.12/site-packages/datasets/packaged_modules/text/text.py", line 98, in _generate_tables
                  batch = f.read(self.config.chunksize)
                          ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.12/site-packages/datasets/utils/file_utils.py", line 844, in read_with_retries
                  out = read(*args, **kwargs)
                        ^^^^^^^^^^^^^^^^^^^^^
                File "<frozen codecs>", line 322, in decode
              UnicodeDecodeError: 'utf-8' codec can't decode byte 0xfe in position 213: invalid start byte

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.

anima agent-lane argument-copy corpus (rung-0)

Byte-level (vocab256) agent-lane tool-USE demo corpus that forces verbatim argument-copy, built to close the key-binding residual of the register-matched tool-use grounding fire (dancinlab/anima-clm-tooluse-rung0-byte-18m, #1833).

Why

The prior agent-lane corpus used only 4 fixed demo keys (ZK7/QX2/MV9/RP4), so a byte-LM could MEMORIZE "vault question -> emit one of those 4 keys" and never had to COPY the asked key. End-to-end grounding stayed at 0/36 because correct_call=0/36.

The fix (this corpus)

Each demo uses a fresh opaque key from a large space (2878 distinct 3-char keys, mean reuse 1.25/key, max 5) so memorization cannot win — only copying the asked key out of the prompt into the call arg generalizes. The call arg is a verbatim copy of the prompt key; the tool-result value is derived deterministically from the key.

Shape (register-matched 사용자:/도우미:, 5-lang, sentinel grammar 0xFE/0xFF):

사용자: <question with a FRESH key K> | 도우미: <reasoning> 0xFE fact_lookup K 0xFF
‹tool-result: fact_lookup K → value(K)›
도우미: grounded — K resolves to value(K).

Shapes a/b/c/d balanced 1200 each (needs-tool / no-tool / don't-know-call / tier-low-refuse).

Invariants

  • bytes=1065956 sha256=ff137ad88b7a4be75e6eaf5126bdc6311b70337100bf91c59a755031e4a1cb32
  • frames 0xFE==0xFF==3600, fabricated_result_count=0, philosophy-grep=0 (no role/persona/system markers — sentinels are learned grammar, not identity; p1..p6)
  • LEAK GUARD: the held-out probe keys PB01..PB36 and their values appear nowhere in the corpus (leak=0, value-collisions=0) — the falsifier stays uncontaminated.
  • deterministic (seed 20260604), reproducible via agent_lane_argcopy_gen.py (included).

Result it produced (honest, a_paper_negative_ok)

A 🔴 CLOSED-NEGATIVE: even with this memorization-proof corpus, the 18M byte-LM did NOT learn verbatim held-out key-binding (correct_call=0/36, grounding=0/36). It generalized the key distribution (emitting shape-plausible keys like P20/LB0), not the copy operation. Ruled-out axis: copy-from-corpus-distribution ⊥ verbatim held-out key-binding at 18M byte scale. Lever moves to an explicit copy-attention / pointer mechanism. Scope: TOY 18M only; mid/7B transfer unverified.
Downloads last month
17
Total file size:
2.8 MB

Collection including dancinlab/anima-agent-lane-argcopy-corpus