Datasets:
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
instruction: string
reference_code: string
output_dict: struct<Q: string, R: string, P: string, L: string, B: string, C: string, E: string, F: string, M: st (... 1239 chars omitted)
child 0, Q: string
child 1, R: string
child 2, P: string
child 3, L: string
child 4, B: string
child 5, C: string
child 6, E: string
child 7, F: string
child 8, M: string
child 9, D: string
child 10, S: string
child 11, G: string
child 12, Diff: string
child 13, A: string
child 14, E1: string
child 15, E2: string
child 16, E3: string
child 17, H: string
child 18, U: string
child 19, T: string
child 20, W: string
child 21, K: string
child 22, Q1: string
child 23, Q2: string
child 24, Q3: string
child 25, R1: string
child 26, R2: string
child 27, R3: string
child 28, S1: string
child 29, S2: string
child 30, S3: string
child 31, P1: string
child 32, P2: string
child 33, P3: string
child 34, V1: string
child 35, V2: string
child 36, V3: string
child 37, T1: string
child 38, T2: string
child 39, T3: string
child 40, U1: string
child 41, U2: string
child 42, U3: string
child 43, W1: string
child 44, W2: string
child 45, W3: string
child 46, X1: string
child 47, X2: string
child 48, X3: string
child 49, Y1: string
child 50, Y2: string
child 51, Y3: string
child 52, A1: string
child 53, A2: string
child 54, A3: string
child 55, V: string
child 56, L1: string
child 57, L2: string
child 58,
...
child 60, Z: string
child 61, M1: string
child 62, M2: string
child 63, M3: string
child 64, E4: string
child 65, N: string
child 66, O: string
child 67, N1: string
child 68, N2: string
child 69, N3: string
child 70, M4: string
child 71, F1: string
child 72, F2: string
child 73, D1: string
child 74, D2: string
child 75, D3: string
child 76, L3: string
child 77, L4: string
child 78, L5: string
child 79, L6: string
child 80, G1: string
child 81, G2: string
child 82, G3: string
child 83, G4: string
child 84, H1: string
child 85, H2: string
child 86, H3: string
child 87, K1: string
child 88, K2: string
child 89, K3: string
child 90, F3: string
child 91, H4: string
child 92, J1: string
child 93, J2: string
child 94, J3: string
child 95, F4: string
child 96, R4: string
child 97, R5: string
child 98, A4: string
child 99, C1: string
child 100, C2: string
child 101, C3: string
child 102, C4: string
child 103, F5: string
child 104, V4: string
child 105, Z1: string
child 106, Z2: string
child 107, Z3: string
child 108, G5: string
child 109, G6: string
child 110, H5: string
child 111, K4: string
child 112, I: string
feature: string
complexity: string
form_stdout: string
messages: list<item: struct<role: string, content: string>>
child 0, item: struct<role: string, content: string>
child 0, role: string
child 1, content: string
form_rc: int64
code: string
task_id: string
to
{'task_id': Value('string'), 'messages': List({'role': Value('string'), 'content': Value('string')}), 'instruction': Value('string'), 'code': Value('string'), 'form_stdout': Value('string'), 'form_rc': 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
instruction: string
reference_code: string
output_dict: struct<Q: string, R: string, P: string, L: string, B: string, C: string, E: string, F: string, M: st (... 1239 chars omitted)
child 0, Q: string
child 1, R: string
child 2, P: string
child 3, L: string
child 4, B: string
child 5, C: string
child 6, E: string
child 7, F: string
child 8, M: string
child 9, D: string
child 10, S: string
child 11, G: string
child 12, Diff: string
child 13, A: string
child 14, E1: string
child 15, E2: string
child 16, E3: string
child 17, H: string
child 18, U: string
child 19, T: string
child 20, W: string
child 21, K: string
child 22, Q1: string
child 23, Q2: string
child 24, Q3: string
child 25, R1: string
child 26, R2: string
child 27, R3: string
child 28, S1: string
child 29, S2: string
child 30, S3: string
child 31, P1: string
child 32, P2: string
child 33, P3: string
child 34, V1: string
child 35, V2: string
child 36, V3: string
child 37, T1: string
child 38, T2: string
child 39, T3: string
child 40, U1: string
child 41, U2: string
child 42, U3: string
child 43, W1: string
child 44, W2: string
child 45, W3: string
child 46, X1: string
child 47, X2: string
child 48, X3: string
child 49, Y1: string
child 50, Y2: string
child 51, Y3: string
child 52, A1: string
child 53, A2: string
child 54, A3: string
child 55, V: string
child 56, L1: string
child 57, L2: string
child 58,
...
child 60, Z: string
child 61, M1: string
child 62, M2: string
child 63, M3: string
child 64, E4: string
child 65, N: string
child 66, O: string
child 67, N1: string
child 68, N2: string
child 69, N3: string
child 70, M4: string
child 71, F1: string
child 72, F2: string
child 73, D1: string
child 74, D2: string
child 75, D3: string
child 76, L3: string
child 77, L4: string
child 78, L5: string
child 79, L6: string
child 80, G1: string
child 81, G2: string
child 82, G3: string
child 83, G4: string
child 84, H1: string
child 85, H2: string
child 86, H3: string
child 87, K1: string
child 88, K2: string
child 89, K3: string
child 90, F3: string
child 91, H4: string
child 92, J1: string
child 93, J2: string
child 94, J3: string
child 95, F4: string
child 96, R4: string
child 97, R5: string
child 98, A4: string
child 99, C1: string
child 100, C2: string
child 101, C3: string
child 102, C4: string
child 103, F5: string
child 104, V4: string
child 105, Z1: string
child 106, Z2: string
child 107, Z3: string
child 108, G5: string
child 109, G6: string
child 110, H5: string
child 111, K4: string
child 112, I: string
feature: string
complexity: string
form_stdout: string
messages: list<item: struct<role: string, content: string>>
child 0, item: struct<role: string, content: string>
child 0, role: string
child 1, content: string
form_rc: int64
code: string
task_id: string
to
{'task_id': Value('string'), 'messages': List({'role': Value('string'), 'content': Value('string')}), 'instruction': Value('string'), 'code': Value('string'), 'form_stdout': Value('string'), 'form_rc': Value('int64')}
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.
FORM Code-Generation Benchmarks
Execution-verified benchmark suite for evaluating LLMs on zero-shot FORM
symbolic-algebra code generation. Every task in every benchmark has been
independently re-validated (instruction clarity, reference/instruction
consistency, output correctness — see benchmarks/adjudication.json).
Revalidation record (v2, September 2026)
Before public release, all four benchmarks underwent a full audit and
re-validation round. The machine-readable adjudication record ships in
benchmarks/adjudication.json.
- 664-det: an automated audit raised 107 flags; manual adjudication (including re-execution of all 664 references under FORM 5.0 with fresh normalized-output comparison) found all 107 to be auditor false positives (e.g. multi-stage pipelines that legitimately print intermediate results). Zero tasks changed; all published numbers for this benchmark are final.
- Open-32: one defective task (id
5905426834, sequential#doderivative loop re-applying its rule to already-modified terms) was fixed with a single correct rule; 3 further flags were judge false positives. The judge is reference-blind, so previously judged model results carry over unchanged; the judge ceiling was re-verified at 32/32 reference solutions judged PASS. - Instruct-100 / Tutorial-44 (v2): an LLM audit flagged 41 of the
original 101 candidate tasks and 21 of 44 tutorial tasks; adjudication
confirmed 31 B2 defects (defective references
replaced with FORM-verified fixes) and 9 B3 defects (dd_ rules placed
before
.sort, wrong wildcard dot patternsv1?.v1?vsv1?.v2?; instructions and references rewritten with verified programs). One misclassified non-executable task was dropped (100 tasks remain; the suite is named Instruct-100 accordingly). Prompt files rebuilt in chat-messages format; all models were re-run on the fixed files: scores shifted (mostly downward) for every model, rankings unchanged. Numbers published after September 2026 refer to the v2 files. - Strict output-match analysis (Tutorial-44 v2): re-executing every
passed program and verifying its result expressions against the reference
(FORM difference-to-zero check) shows that only a minority of runnable
programs genuinely solve the task (per model: 32–100% of passes). The
per-task adjudication data is in
benchmarks/true_solve_adjudication.json. - Protocol notes: frontier baselines evaluated single-attempt with a 6,000-character syntax guide in context (temp 0); fine-tuned models zero-shot without docs. Local GPU-node runs were verified with FORM 4.3; laptop/cloud runs with FORM 5.0.
Files
| File | Tasks | Metric | Description |
|---|---|---|---|
benchmarks/v3/test.jsonl |
664 | exact output match | Deterministic tasks with expected FORM output (from reference execution) |
benchmarks/splits/test.jsonl |
100 | syntax pass; strict output match (89 adjudicable, lower bound) | Open-ended instructions with reference programs (v2, re-validated) |
benchmarks/tutorial/tutorial_test_v2.jsonl |
44 | syntax pass; strict output match (41 adjudicable) | Tasks derived from FORM tutorial programs, rewritten as plain-English instructions (v2: re-validated) |
benchmarks/v3c/test_open_v2.jsonl |
32 | rubric (anchored judge); strict output match (22 output-determined) | Open-ended generation tasks with reference programs |
benchmarks/v3c/test_open_v2_docs.jsonl |
32 | rubric (anchored judge) | Same tasks, with FORM docs prepended (for docs-in-context baselines) |
benchmarks/per_task_verdicts.json |
— | — | Per-task verdicts (exec / rubric / strict) for all evaluated models on all four benchmarks; every aggregate verified against the paper's Table 3 |
Format
All files are JSONL. Two schemas:
- Deterministic (
v3/test.jsonl):{instruction, reference_code, expected_output, feature, complexity, form_stdout} - Chat-style (
splits/test.jsonl,tutorial_test_v2.jsonl):{messages: [{role: user, content: instruction}, {role: assistant, content: reference program}]}
Training data: the training corpus is not part of this release; it is available from the author upon request.
Verification protocol
- Every reference program runs through FORM 5.0 with exit code 0 and deterministic output (verifiable by re-running).
- 664-det
expected_outputvalues were computed by executing the reference through FORM and normalizing the printed expressions. - Instructions contain no FORM syntax (checked programmatically) so the benchmark measures zero-shot code generation, not instruction following with scaffolding.
Evaluation
- Output match: run generated program; compare per-expression normalized
output against
expected_output. - Syntax pass: program runs through FORM (exit 0, non-empty output, no errors).
- Rubric (open-32): anchored judge scores each generated program against per-task requirements extracted from the instruction (PASS iff score = 1.0).
License
CC-BY-4.0. Reference programs are original or paraphrase-derived; no verbatim third-party documentation is included in the test sets.
- Downloads last month
- 22