Motherlode
Collection
A 33.4M-parameter code-search encoder fine-tuned from bge-small-en-v1.5, the evidence behind its figures, and a Space that runs it in a browser. • 3 items • Updated
Dataset Preview
The full dataset viewer is not available (click to read why). Only showing a preview of the rows.
The dataset generation failed
Error code: DatasetGenerationError
Exception: CastError
Message: Couldn't cast
site: string
local_models: bool
chrome: string
gpu: null
webgpu: struct<path: string, ready_ms: int64, reader: string, memReader: string, memMean: string, rows: int6 (... 148 chars omitted)
child 0, path: string
child 1, ready_ms: int64
child 2, reader: string
child 3, memReader: string
child 4, memMean: string
child 5, rows: int64
child 6, first: struct<key: string, score: double, text: string>
child 0, key: string
child 1, score: double
child 2, text: string
child 7, planted_first: bool
child 8, hits: list<item: struct<key: string, score: double, text: string>>
child 0, item: struct<key: string, score: double, text: string>
child 0, key: string
child 1, score: double
child 2, text: string
wasm: struct<path: string, ready_ms: int64, reader: string, memReader: null, memMean: null, rows: int64, f (... 144 chars omitted)
child 0, path: string
child 1, ready_ms: int64
child 2, reader: string
child 3, memReader: null
child 4, memMean: null
child 5, rows: int64
child 6, first: struct<key: string, score: double, text: string>
child 0, key: string
child 1, score: double
child 2, text: string
child 7, planted_first: bool
child 8, hits: list<item: struct<key: string, score: double, text: string>>
child 0, item: struct<key: string, score: double, text: string>
child 0, key: string
child 1, score: double
child 2, text: string
d1_passes: bool
console_tail: list<item: string>
child 0, item: string
deployed: string
mirror_requests: string
reader_files: string
onnx_sha256: string
onnx_bytes: int64
texts: int64
opset: int64
min_cosine: double
passes: bool
prose: int64
torch: string
onnxruntime: string
max_abs_diff: double
code: int64
soup_safetensors_sha256: string
to
{'texts': Value('int64'), 'code': Value('int64'), 'prose': Value('int64'), 'min_cosine': Value('float64'), 'max_abs_diff': Value('float64'), 'passes': Value('bool'), 'opset': Value('int64'), 'torch': Value('string'), 'onnxruntime': Value('string'), 'onnx_sha256': Value('string'), 'onnx_bytes': Value('int64'), 'soup_safetensors_sha256': Value('string')}
because column names don't match
Traceback: Traceback (most recent call last):
File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1827, in _prepare_split_single
for key, table in generator:
^^^^^^^^^
File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 613, in wrapped
for item in generator(*args, **kwargs):
~~~~~~~~~^^^^^^^^^^^^^^^^^
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
site: string
local_models: bool
chrome: string
gpu: null
webgpu: struct<path: string, ready_ms: int64, reader: string, memReader: string, memMean: string, rows: int6 (... 148 chars omitted)
child 0, path: string
child 1, ready_ms: int64
child 2, reader: string
child 3, memReader: string
child 4, memMean: string
child 5, rows: int64
child 6, first: struct<key: string, score: double, text: string>
child 0, key: string
child 1, score: double
child 2, text: string
child 7, planted_first: bool
child 8, hits: list<item: struct<key: string, score: double, text: string>>
child 0, item: struct<key: string, score: double, text: string>
child 0, key: string
child 1, score: double
child 2, text: string
wasm: struct<path: string, ready_ms: int64, reader: string, memReader: null, memMean: null, rows: int64, f (... 144 chars omitted)
child 0, path: string
child 1, ready_ms: int64
child 2, reader: string
child 3, memReader: null
child 4, memMean: null
child 5, rows: int64
child 6, first: struct<key: string, score: double, text: string>
child 0, key: string
child 1, score: double
child 2, text: string
child 7, planted_first: bool
child 8, hits: list<item: struct<key: string, score: double, text: string>>
child 0, item: struct<key: string, score: double, text: string>
child 0, key: string
child 1, score: double
child 2, text: string
d1_passes: bool
console_tail: list<item: string>
child 0, item: string
deployed: string
mirror_requests: string
reader_files: string
onnx_sha256: string
onnx_bytes: int64
texts: int64
opset: int64
min_cosine: double
passes: bool
prose: int64
torch: string
onnxruntime: string
max_abs_diff: double
code: int64
soup_safetensors_sha256: string
to
{'texts': Value('int64'), 'code': Value('int64'), 'prose': Value('int64'), 'min_cosine': Value('float64'), 'max_abs_diff': Value('float64'), 'passes': Value('bool'), 'opset': Value('int64'), 'torch': Value('string'), 'onnxruntime': Value('string'), 'onnx_sha256': Value('string'), 'onnx_bytes': Value('int64'), 'soup_safetensors_sha256': Value('string')}
because column names don't match
The above exception was the direct cause of the following exception:
Traceback (most recent call last):
File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 1369, in compute_config_parquet_and_info_response
parquet_operations, partial, estimated_dataset_info = stream_convert_to_parquet(
~~~~~~~~~~~~~~~~~~~~~~~~~^
builder, max_dataset_size_bytes=max_dataset_size_bytes
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
)
^
File "/src/services/worker/src/worker/job_runners/config/parquet_and_info.py", line 948, in stream_convert_to_parquet
builder._prepare_split(split_generator=splits_generators[split], file_format="parquet")
~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1694, in _prepare_split
for job_id, done, content in self._prepare_split_single(
~~~~~~~~~~~~~~~~~~~~~~~~~~^
gen_kwargs=gen_kwargs, job_id=job_id, **_prepare_split_args
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
):
^
File "/usr/local/lib/python3.14/site-packages/datasets/builder.py", line 1880, in _prepare_split_single
raise DatasetGenerationError("An error occurred while generating the dataset") from e
datasets.exceptions.DatasetGenerationError: An error occurred while generating the datasetNeed help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
texts int64 | code int64 | prose int64 | min_cosine float64 | max_abs_diff float64 | passes bool | opset int64 | torch string | onnxruntime string | onnx_sha256 string | onnx_bytes int64 | soup_safetensors_sha256 string |
|---|---|---|---|---|---|---|---|---|---|---|---|
2,000 | 1,000 | 1,000 | 0.842787 | 0.211957 | false | 17 | 2.8.0 | 1.23.2 | 21b8550df9c72321f492a8d6045bd4074a785aa424238abdeecc6676735868ef | 133,029,886 | 637a02871f785c69aa54ea6ae7e543692b1e6ce3596347739400ba43fa32b8e5 |
Evidence for motherlode-code-small-en-v0.1
Every figure on the card of RiverRider/motherlode-code-small-en-v0.1
comes from a file here. Each figure also has a row in Sunstone North Labs' claims ledger, named in the first column. The
table maps each row to its file and field. The per-instance and per-query files let you recompute a paired test without
our code or our hardware.
"The soup" is the model: the parameter-wise mean of two fine-tunes of bge-small-en-v1.5, a2 and d1, at weight 0.5.
"The 411" are the SWE-bench Verified instances left after 89 were flagged for a gold file sharing text with d1's
training texts. The flagged list is in evidence/tier3b_2026-09-26/swe/swe_overlap_d1.json, key flagged.
Where each figure is
| row | figure | file | field |
|---|---|---|---|
| L1 | finding files through gate 1's base pipeline: 178 of the 411 against bge-small's 145, 219 of 500 against 181, and the peers | evidence/soup_peers_2026-09-27/summary.json |
swe.recall1_unflagged, swe.counts, swe.test1_soup_vs_bge_small_unflagged |
| L1 | the same, per instance and model | evidence/soup_peers_2026-09-27/swe/<model>-<django,rest,sympy>/instances.jsonl |
rank of the gold file, 1 is a hit |
| L2 | a replay of the engine's published path, each model with its shipped mean: 230 of the 411 against 192 | evidence/soup_integration_2026-09-27/M1_code.json |
p0 |
| L2 | the same, per instance: the soup and bge-small | evidence/soup_integration_2026-09-27/soup-mixed-p0.jsonl, evidence/engine_replay_2026-09-27/bge-small-shipped-p0.jsonl |
rank |
| L2 | each model with its own code mean, the reproduction of the published run, and the floors | evidence/engine_replay_2026-09-27/summary.json |
test_soup_vs_bge_small_pilot_means_p0_unflagged, reproduction, floor_recall1 |
| L2 | the engine at 8234be7 |
evidence/soup_integration_2026-09-27/M1_code.json, and the -p1.jsonl files |
p1 |
| L3 | CoIR, the soup | evidence/soup_peers_2026-09-27/summary.json, evidence/soup_peers_2026-09-27/coir/soup/*.perquery.jsonl |
coir |
| L3 | CoIR, the other models, per query | evidence/tier3_2026-09-26/perquery.tar.xz, evidence/tier3b_2026-09-26/perquery.tar.xz |
qid, ndcg10 |
| L4 | scifact and COCO under the shipped means | evidence/soup_integration_2026-09-27/M1_prose.json |
M1_prose |
| L4 | the same with the soup's code mean, and the shuffled floors | evidence/soup_engine_2026-09-27/I4_prose.json |
scifact, coco |
| L5 | bge-small's shipped mean and the COCO test captions | evidence/soup_integration_2026-09-27/M1_bge_mean.json |
every field |
| L6 | ONNX, GGUF and wasm reader parity | evidence/soup_engine_2026-09-27/I2_export.json, evidence/soup_integration_2026-09-27/G1.json, W1.json |
min_cosine |
| L7 | transformers.js and the engine's keep-[SEP] call | evidence/soup_engine_2026-09-27/I3_tfjs.json, evidence/soup_integration_2026-09-27/I3b.json |
min_cosine |
| L8 | anisotropy | evidence/engine_replay_2026-09-27/summary.json, evidence/soup_integration_2026-09-27/M1_code.json, M1_prose.json, evidence/soup_engine_2026-09-27/I4_prose.json |
arms, rows_after_shipped_mean, anisotropy_after_mean, anisotropy |
| L9 | the engine's gate constants | evidence/soup_integration_2026-09-27/I5_gates_shipped.json |
floors, band |
| L10 | contamination | evidence/tier3b_2026-09-26/swe/swe_overlap_d1.json, evidence/tier3b_2026-09-26/peers/overlap_d1.json, evidence/tier3_2026-09-26/overlap.json |
flagged |
| L11 | the three refitted heads | evidence/soup_integration_2026-09-27/H1.json, H2.json, H3.json |
soup_minus_bge_small, control_holds |
| L12 | the wasm reader's 64-token cut | evidence/soup_integration_2026-09-27/W1.json |
the 64-token cut against 512 tokens |
| L13 | retrieval on the wasm path, at 64, 256 and 512 tokens, and its control | evidence/wasm_path_2026-09-27/summary.json, control.json, speed.json |
W2a_*, W2b_*, rows |
| L13 | the same, per instance | evidence/wasm_path_2026-09-27/<reader>-<cut>-<p0,p1>.jsonl |
rank |
| L14 | the Mac app end to end: recall on real prose, and pictures | evidence/app_e2e_2026-09-27/A1_recall.json, A1_questions.json, A2_images.json |
A1_soup_vs_bge_small_dense_rank1, rows, summary_line |
| L15 | the release: the model loaded from the Hub by its id without a token, by sentence-transformers 5.1.0 and 6.0.1 and by transformers.js 3.7.1 with the engine's call; every file's hash; the GGUF's tensors against the file G1 tested; the wasm reader at 256 tokens; the Mac app and Sunstone loading it by name | evidence/release_2026-09-27/hub_load.jsonl, gguf_reconvert.json, W1b_256.json, app_hub.json, sunstone_hub.json |
lines 10 to 13 of hub_load.jsonl, tensor_data_equal, min_cosine, equals_tested_gguf, checks_with_ok |
| L16 | video through the app's rows, MSR-VTT's 1k-A test split: 0.2231 against 0.2129 at recall@1, 20,000 captions over 1,000 videos | evidence/followups_2026-09-27/V1_video.json |
test, readers, controls |
| L16 | the same, per caption: the rank of the first row of the caption's video in each store | evidence/followups_2026-09-27/V1_ranks.jsonl |
soup, bge-small |
| L17 | the half-precision export: parity, I4 through it, and the page on WebGPU | evidence/followups_2026-09-27/F1_parity.json, F1_I4.json, F1b_page.json |
min_cosine, fp16_minus_float32, dtypes |
| L18 | the engine's wasm at opt-level 3 and z, and the build with +simd128 that does not compile |
evidence/followups_2026-09-27/S1_wasm.json, S1b_engine_sizes.json, S1_simd128_build.json |
variants, opt_level_3, errors |
| L19 | the default reader on the page, locally, on the dev copy and live, and in Sunstone, with the upgrade from 0.1.3 | evidence/followups_2026-09-27/D1_page_local.json, D1_page_dev.json, D1_page_live.json, D1_sunstone_rate.json, D1_sunstone_upgrade.json |
d1_passes, runs, opening_restore |
| L20 | the default reader in the Mac app: a bge-small store carried, and A1 on it | evidence/followups_2026-09-27/D1_app.json |
carry, rank1 |
| L22 | the page's float32 batches on WebGPU: a batch's last row returned non-finite, and the appended row that stops it | evidence/followups_2026-09-27/F1c_webgpu_nan.json |
what_it_follows, fix, fix_through_embedKeepSep_against_pytorch, page_end_to_end |
| L23 | the engine's tool path without its lexical bonus, the soup with its shipped mean: 221 of the 411 against 207 with the bonus, 272 of 500 against 255, and every other cell | evidence/engine_nobonus_2026-09-28/summary.json |
arms.soup-mixed: removed_alone_from_head_unflagged, removed_alone_from_head, recall1_unflagged, recall1 |
| L26 | the tool path as it ships from engine a3738ff, on blackwindow.xyz and in Sunstone 0.1.6: 272 of 500 against bge-small's 236, and 221 of the 411 against 194 |
evidence/engine_nobonus_2026-09-28/summary.json |
soup_mixed_vs_bge_small_shipped, under HEAD without L |
| L23, L26 | the same, per instance, for the soup and for bge-small | evidence/engine_nobonus_2026-09-28/soup-mixed.jsonl, evidence/engine_decompose_2026-09-27/bge-small-shipped.jsonl |
ranks, keyed by the changes a cell keeps, as parts in the summary names them: HEAD is RLKTD, HEAD without the bonus RKTD, the published path - |
| L37 | the lexical bonus by reader: the cells and moves of a re-pairing from the two files above, the bonus added after centring, and the bonus in units of each query's own top-10 spread, registered before its run | evidence/engine_bonus_scale_2026-09-29/repair.json, summary.json, and per instance soup-mixed.jsonl and bge-small-shipped.jsonl |
cells, bonus_within_reader, median_raw_spread_top10_hits_p1; arms.*.scaled_vs_none, motherlode_vs_bge_small_unflagged, comes_back; per instance ranks under RKTD, RLKTD and RSKTD, the scaled bonus |
hub_load.jsonl keeps every run in order. Lines 1 to 5 ran while the repository was private, with the account's token,
and line 3 fails on the check's own defect, since a hash list cannot list itself. Lines 6 to 9 ran after publication with
no token, and record anonymous false because the first version of the flag counted HF_HUB_DISABLE_IMPLICIT_TOKEN
as a token. Lines 10 to 13 are the public check. The check of this dataset's own files against its SHA256SUMS is
not among them: a file that counts the dataset's files cannot be one of them, so it stays in the lodestone repository.
Every paired test on SWE-bench is recall@1 paired by instance, with an exact two-sided sign test over the instances only
one model gets right. The prose and head comparisons use 2,000 bootstrap resamples, paired by query or caption. Their
per-query values are not stored. gates/soup_checks.py and gates/soup_heads.py recompute them from public data and
the model.
The rest of the folder
evidence/engine_decompose_2026-09-27/: the engine's HEAD regression split into five changes, 32 cells a model.evidence/soup_engine_2026-09-27/I1_identity.json: the soup's configuration and tokenizer against bge-small's.evidence/soup_engine_2026-09-27/I2_export_first_run_train_mode.json: the first run of the export check. It failed because its reference model was left in training mode, and it is kept because it failed.evidence/soup_engine_2026-09-27/I5_gates.json: the gate constants under the code mean, before the shipped mean.pilot/repos.txt: the 86 Python repositories a2's corpus came from, each with the licence read from its own file.gates/: the measurement scripts that wrote these files. The training code is not included. Paths under/root/and/tmp/in the files are the rented box's and scratch working directories.SHA256SUMS: every file's hash.
Licence and citation
Artifacts and scripts are under CC BY 4.0. The instances derive from SWE-bench Verified and the code-search queries from CoIR, whose licences and citations apply to the underlying data. The model these files measure is licensed separately, under the Business Source License 1.1 in its own repository.
Sunstone North Labs. Contact: burton@sunstonenorth.com
- Downloads last month
- 779