Hugging Face's logo

Datasets:
pragnyanramtha
/
splice-contextbench

ArXiv:
Tags:
contextbench
swe-bench
code-agents
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
instance_id: string
traj_data: struct<pred_steps: list<item: struct<files: list<item: string>, spans: struct<django/contrib/admin/f (... 1266 chars omitted)
  child 0, pred_steps: list<item: struct<files: list<item: string>, spans: struct<django/contrib/admin/filters.py: list<ite (... 1028 chars omitted)
      child 0, item: struct<files: list<item: string>, spans: struct<django/contrib/admin/filters.py: list<item: struct<t (... 1016 chars omitted)
          child 0, files: list<item: string>
              child 0, item: string
          child 1, spans: struct<django/contrib/admin/filters.py: list<item: struct<type: string, start: int64, end: int64>>,  (... 974 chars omitted)
              child 0, django/contrib/admin/filters.py: list<item: struct<type: string, start: int64, end: int64>>
                  child 0, item: struct<type: string, start: int64, end: int64>
                      child 0, type: string
                      child 1, start: int64
                      child 2, end: int64
              child 1, django/db/models/fields/related.py: list<item: struct<type: string, start: int64, end: int64>>
                  child 0, item: struct<type: string, start: int64, end: int64>
                      child 0, type: string
                      child 1, start: int64
                      child 2, end: int64
              child 2, django/db/models/fields/__init__.py: list<item: struct<type: string, start: int64, end: int64>>
                  child 0, item: struct<typ
...
ruct<recall: double, precision: double, intersection: int64, gold_size: int64, pred_size: int64>
  child 0, recall: double
  child 1, precision: double
  child 2, intersection: int64
  child 3, gold_size: int64
  child 4, pred_size: int64
final: struct<file: struct<coverage: double, precision: double, intersection: int64, gold_size: int64, pred (... 340 chars omitted)
  child 0, file: struct<coverage: double, precision: double, intersection: int64, gold_size: int64, pred_size: int64>
      child 0, coverage: double
      child 1, precision: double
      child 2, intersection: int64
      child 3, gold_size: int64
      child 4, pred_size: int64
  child 1, symbol: struct<coverage: double, precision: double, intersection: int64, gold_size: int64, pred_size: int64>
      child 0, coverage: double
      child 1, precision: double
      child 2, intersection: int64
      child 3, gold_size: int64
      child 4, pred_size: int64
  child 2, span: struct<coverage: double, precision: double, intersection: int64, gold_size: int64, pred_size: int64>
      child 0, coverage: double
      child 1, precision: double
      child 2, intersection: int64
      child 3, gold_size: int64
      child 4, pred_size: int64
  child 3, line: struct<coverage: double, precision: double, intersection: int64, gold_size: int64, pred_size: int64>
      child 0, coverage: double
      child 1, precision: double
      child 2, intersection: int64
      child 3, gold_size: int64
      child 4, pred_size: int64
to
{'instance_id': Value('string'), 'num_steps': Value('int64'), 'final': {'file': {'coverage': Value('float64'), 'precision': Value('float64'), 'intersection': Value('int64'), 'gold_size': Value('int64'), 'pred_size': Value('int64')}, 'symbol': {'coverage': Value('float64'), 'precision': Value('float64'), 'intersection': Value('int64'), 'gold_size': Value('int64'), 'pred_size': Value('int64')}, 'span': {'coverage': Value('float64'), 'precision': Value('float64'), 'intersection': Value('int64'), 'gold_size': Value('int64'), 'pred_size': Value('int64')}, 'line': {'coverage': Value('float64'), 'precision': Value('float64'), 'intersection': Value('int64'), 'gold_size': Value('int64'), 'pred_size': Value('int64')}}, 'trajectory': {'steps': List({'step': Value('int64'), 'coverage': {'file': Value('float64'), 'symbol': Value('float64'), 'span': Value('float64'), 'line': Value('float64')}}), 'auc_coverage': {'file': Value('float64'), 'symbol': Value('float64'), 'span': Value('float64'), 'line': Value('float64')}, 'redundancy': {'file': Value('float64'), 'symbol': Value('float64'), 'span': Value('float64'), 'line': Value('float64')}}, 'editloc': {'recall': Value('float64'), 'precision': Value('float64'), 'intersection': Value('int64'), 'gold_size': Value('int64'), 'pred_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
              instance_id: string
              traj_data: struct<pred_steps: list<item: struct<files: list<item: string>, spans: struct<django/contrib/admin/f (... 1266 chars omitted)
                child 0, pred_steps: list<item: struct<files: list<item: string>, spans: struct<django/contrib/admin/filters.py: list<ite (... 1028 chars omitted)
                    child 0, item: struct<files: list<item: string>, spans: struct<django/contrib/admin/filters.py: list<item: struct<t (... 1016 chars omitted)
                        child 0, files: list<item: string>
                            child 0, item: string
                        child 1, spans: struct<django/contrib/admin/filters.py: list<item: struct<type: string, start: int64, end: int64>>,  (... 974 chars omitted)
                            child 0, django/contrib/admin/filters.py: list<item: struct<type: string, start: int64, end: int64>>
                                child 0, item: struct<type: string, start: int64, end: int64>
                                    child 0, type: string
                                    child 1, start: int64
                                    child 2, end: int64
                            child 1, django/db/models/fields/related.py: list<item: struct<type: string, start: int64, end: int64>>
                                child 0, item: struct<type: string, start: int64, end: int64>
                                    child 0, type: string
                                    child 1, start: int64
                                    child 2, end: int64
                            child 2, django/db/models/fields/__init__.py: list<item: struct<type: string, start: int64, end: int64>>
                                child 0, item: struct<typ
              ...
              ruct<recall: double, precision: double, intersection: int64, gold_size: int64, pred_size: int64>
                child 0, recall: double
                child 1, precision: double
                child 2, intersection: int64
                child 3, gold_size: int64
                child 4, pred_size: int64
              final: struct<file: struct<coverage: double, precision: double, intersection: int64, gold_size: int64, pred (... 340 chars omitted)
                child 0, file: struct<coverage: double, precision: double, intersection: int64, gold_size: int64, pred_size: int64>
                    child 0, coverage: double
                    child 1, precision: double
                    child 2, intersection: int64
                    child 3, gold_size: int64
                    child 4, pred_size: int64
                child 1, symbol: struct<coverage: double, precision: double, intersection: int64, gold_size: int64, pred_size: int64>
                    child 0, coverage: double
                    child 1, precision: double
                    child 2, intersection: int64
                    child 3, gold_size: int64
                    child 4, pred_size: int64
                child 2, span: struct<coverage: double, precision: double, intersection: int64, gold_size: int64, pred_size: int64>
                    child 0, coverage: double
                    child 1, precision: double
                    child 2, intersection: int64
                    child 3, gold_size: int64
                    child 4, pred_size: int64
                child 3, line: struct<coverage: double, precision: double, intersection: int64, gold_size: int64, pred_size: int64>
                    child 0, coverage: double
                    child 1, precision: double
                    child 2, intersection: int64
                    child 3, gold_size: int64
                    child 4, pred_size: int64
              to
              {'instance_id': Value('string'), 'num_steps': Value('int64'), 'final': {'file': {'coverage': Value('float64'), 'precision': Value('float64'), 'intersection': Value('int64'), 'gold_size': Value('int64'), 'pred_size': Value('int64')}, 'symbol': {'coverage': Value('float64'), 'precision': Value('float64'), 'intersection': Value('int64'), 'gold_size': Value('int64'), 'pred_size': Value('int64')}, 'span': {'coverage': Value('float64'), 'precision': Value('float64'), 'intersection': Value('int64'), 'gold_size': Value('int64'), 'pred_size': Value('int64')}, 'line': {'coverage': Value('float64'), 'precision': Value('float64'), 'intersection': Value('int64'), 'gold_size': Value('int64'), 'pred_size': Value('int64')}}, 'trajectory': {'steps': List({'step': Value('int64'), 'coverage': {'file': Value('float64'), 'symbol': Value('float64'), 'span': Value('float64'), 'line': Value('float64')}}), 'auc_coverage': {'file': Value('float64'), 'symbol': Value('float64'), 'span': Value('float64'), 'line': Value('float64')}, 'redundancy': {'file': Value('float64'), 'symbol': Value('float64'), 'span': Value('float64'), 'line': Value('float64')}}, 'editloc': {'recall': Value('float64'), 'precision': Value('float64'), 'intersection': Value('int64'), 'gold_size': Value('int64'), 'pred_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.

Splice on ContextBench

Splice is a context-retrieval harness for ContextBench. It reuses the SWE-bench Docker loop (not a from-scratch agent) and publishes as Splice, not mini-SWE.

This dataset is a budget slice, not the 1,136-task full board. Goal: benchmark the method as far as budget allows, not claim #1 vs 57.5% Pass@1. We run historic-fail / multi-file Verified tasks one-by-one. Instances mini-SWE already clears are deprioritized.

Field Value
Harness Splice (working-set <PATCH_CONTEXT>; spans >40 lines are dropped from that set)
Backbone DeepSeek-V4-Pro-0813 via Fireworks
Slice Historic-fail / localization-hard SWE-bench Verified
Live board (full set, not us) contextbench.github.io

Board (this slice)

Numbers below are filled as tasks finish. Line F1 is the harmonic mean of line-level coverage and precision vs gold spans. Submitted ≠ Pass@1 (official tests not yet scored on every row). Published ContextBench Backbone Only (full set, mini-SWE + DeepSeek-V4-Pro): 57.5% Pass@1, Line F1 0.338. That row is a different snapshot and instance set — do not read this slice F1 as a full-board claim.

Instance Historic Splice exit $ Line cov Line prec Line F1 Pass@1
scikit-learn-25232 pass Submitted 0.25 0.943 0.125 0.221 yes (1/1 passed)
django-16263 fail RuntimeError 2.00 — — — no (no submitted patch ($2 cap))
sympy-16597 fail Submitted 0.90 0.000 0.000 0.000 no (0/3 F2P (assumptions tests))
matplotlib-14623 fail Submitted 0.20 0.000 0.000 0.000 yes (1/1 passed)
django-11400 fail Submitted 0.17 0.000 0.000 0.000 no (4/6 F2P (get_choices ordering))

What Splice changes

  1. Working set — every explicit sed -n / nl|sed / head / <EXPLORE_CONTEXT> span is recorded. On submit, Splice emits <PATCH_CONTEXT> from that set (usage-drop fix). Spans wider than 40 lines and test files are not stored.
  2. No execute rewrite — the shell sees the model's command as-is. Rewriting wide sed/head before execute looped (65+ steps, never submitted) on scikit-learn-25232.
  3. Limit abort — if the cost/step cap fires, Splice still writes <PATCH_CONTEXT> from the working set and captures git diff so Line F1 and Pass@1 remain defined. (django-16263 ran before this.)
  4. Name — board row is Splice + DeepSeek-V4-Pro-0813.

Where it should be better

Failure (paper Appendix I) Splice response
Usage drop (saw gold, dropped it before the patch) Harness-owned working set → <PATCH_CONTEXT>
Over-recall (whole-file cat) Wide reads still execute; they are omitted from <PATCH_CONTEXT>
Wrong-file localization Still model-led grep; we prioritize historic-fail multi-file tasks

v0 on 25232 already hit all gold files (file coverage 1.0) but Line precision was 0.125 because the model dumped 1–700 line ranges. Remaining budget goes to historic-fail tasks, not clamped reruns of historic-pass instances.

django-16263 (historic fail) hit the $2 global cost cap at 126 steps without submit: annotations tests were passing, but there was no <PATCH_CONTEXT>, so Line F1 is unscored and Pass@1 is no. Later one-by-one runs use the yaml $3 cap (MSWEA_GLOBAL_COST_LIMIT=3).

Files

  • board.json — per-instance metrics
  • preds/ — patches
  • traj/ — mini-SWE-format trajectories
  • eval/ — ContextBench evaluate JSONL

Reproduce 

export FIREWORKS_API_KEY=...
scripts/run_one.sh scikit-learn__scikit-learn-25232

Model id: fireworks_ai/accounts/fireworks/models/deepseek-v4-pro-0813 (not preview deepseek-v4-pro).
Downloads last month
68
Total file size:
3.7 MB

Paper for pragnyanramtha/splice-contextbench