id stringlengths 64 64 | row_json stringlengths 3.06k 72.9k |
|---|---|
000b55130eae41dbc163f68a664f74266d300745a645694a63373c2c82cfe06e | {"version":"raw-text-cfr-source-native-structure-0.1","track":"regulations","prompt_variant":"cfr-native-lexical-sparse-v2-regulations","document_id":"000b55130eae41dbc163f68a664f74266d300745a645694a63373c2c82cfe06e","split_group_id":"3c798134388c7e6a405048bf4319b5b05ef961000ab6f165a70d819b11fef646","cfr_identifier":"/... |
005a17510d63640aff24dfef6447ca4dadd4b651774a9a2e641ac41c747f6b93 | {"version":"raw-text-cfr-source-native-structure-0.1","track":"regulations","prompt_variant":"cfr-native-lexical-sparse-v2-regulations","document_id":"005a17510d63640aff24dfef6447ca4dadd4b651774a9a2e641ac41c747f6b93","split_group_id":"2ce23791e20ebe163830264e3f4d6453c4c377f282c7b99ecadb2495b6bc3004","cfr_identifier":"/... |
005e8c99fa7a0999030c049d96852eac735639a7e094208d3645682db58e5e18 | {"version":"raw-text-cfr-source-native-structure-0.1","track":"regulations","prompt_variant":"cfr-native-lexical-sparse-v2-regulations","document_id":"005e8c99fa7a0999030c049d96852eac735639a7e094208d3645682db58e5e18","split_group_id":"3955354ab670f0573397130de4983dfc90b629e089bfa84cab5322588f63156b","cfr_identifier":"/... |
006cc518820ec29cb4f89ca5354d356028bea8592ea2f382a09eaa1fe309f261 | {"version":"raw-text-cfr-source-native-structure-0.1","track":"regulations","prompt_variant":"cfr-native-lexical-sparse-v2-regulations","document_id":"006cc518820ec29cb4f89ca5354d356028bea8592ea2f382a09eaa1fe309f261","split_group_id":"2dfdf20600f78632d98f5a8f542d5b46c4fc74312b0c707359ef61abacd7f5db","cfr_identifier":"/... |
00a011932807da2ec90f873047cca77658a0ee85ad5f501ca5759d30e6c62782 | {"version":"raw-text-cfr-source-native-structure-0.1","track":"regulations","prompt_variant":"cfr-native-lexical-sparse-v2-regulations","document_id":"00a011932807da2ec90f873047cca77658a0ee85ad5f501ca5759d30e6c62782","split_group_id":"cc1a66989d9bf426cee8d1217f40fb6ed1b2796c4c202178f563199111e98c04","cfr_identifier":"/... |
00a7fad704f5ae11816fd2aacc13384dfdb0071273c02ff7b0c3fff12e736c47 | {"version":"raw-text-cfr-source-native-structure-0.1","track":"regulations","prompt_variant":"cfr-native-lexical-sparse-v2-regulations","document_id":"00a7fad704f5ae11816fd2aacc13384dfdb0071273c02ff7b0c3fff12e736c47","split_group_id":"a8d39fa74727ffe2222d35b756c6cfdac8303c61a1f47417bc93f3635f214bfb","cfr_identifier":"/... |
00ea2b2309fa04d9177ddec7a19ff2b35384c37449fa69785015d1ac264cdd28 | {"version":"raw-text-cfr-source-native-structure-0.1","track":"regulations","prompt_variant":"cfr-native-lexical-sparse-v2-regulations","document_id":"00ea2b2309fa04d9177ddec7a19ff2b35384c37449fa69785015d1ac264cdd28","split_group_id":"9a2c4c52e5b3d1908f87d786f002ee151d72ad1a08e27b80476aa14318d3b274","cfr_identifier":"/... |
00f5f9074407d0c5078c39bca5f86652d0654c5ac42e2201a7ff82bc8d827cf2 | {"version":"raw-text-cfr-source-native-structure-0.1","track":"regulations","prompt_variant":"cfr-native-lexical-sparse-v2-regulations","document_id":"00f5f9074407d0c5078c39bca5f86652d0654c5ac42e2201a7ff82bc8d827cf2","split_group_id":"e59008dd59fe439b5bef82d785021d7ca366bba2ef583318c3774ef84bef76ba","cfr_identifier":"/... |
011c5ce72879a6a2f40e19480a0b5645bf8168b3c7a65aef7dc4093a80b792f0 | {"version":"raw-text-cfr-source-native-structure-0.1","track":"regulations","prompt_variant":"cfr-native-lexical-sparse-v2-regulations","document_id":"011c5ce72879a6a2f40e19480a0b5645bf8168b3c7a65aef7dc4093a80b792f0","split_group_id":"6d2b26fd0bedcc7e2675945dace49f628aa93307edd7467b86c1445955d5d7d4","cfr_identifier":"/... |
0126ade0e248d554b38edead400423c23d41532bc896c11e9588fb5a3b2af0b5 | {"version":"raw-text-cfr-source-native-structure-0.1","track":"regulations","prompt_variant":"cfr-native-lexical-sparse-v2-regulations","document_id":"0126ade0e248d554b38edead400423c23d41532bc896c11e9588fb5a3b2af0b5","split_group_id":"07ec78a9171b648db5059fedff0ba4658501c21cac8330557a99740e31176319","cfr_identifier":"/... |
013ce54d5c343a2c2188800d4e9a511e94c9e2b44d6aef2e8e130c62cfd85510 | {"version":"raw-text-cfr-source-native-structure-0.1","track":"regulations","prompt_variant":"cfr-native-lexical-sparse-v2-regulations","document_id":"013ce54d5c343a2c2188800d4e9a511e94c9e2b44d6aef2e8e130c62cfd85510","split_group_id":"a8a21665d71424bd7aeef95a4399290ff73a11305cbe71e430efa50eb9b5653e","cfr_identifier":"/... |
XLL Regulations Qwen3.5-9B — Training Set and Evaluation Benchmarks
Replication package for
xll-regulations-qwen3.5-9b
(full-weight fine-tune of Qwen/Qwen3.5-9B for eCFR text → source-native XLL
structure conversion, protocol raw-text-cfr-source-native-structure-0.2,
prompt cfr-native-lexical-sparse-v2-regulations).
Training and directly attested evaluation hashes are checked against the published
model reports. Other provenance files are preserved unchanged; see
reproduction/hash-binding.json for the exact scope of verification.
Files
| File | Size | Attested by |
|---|---|---|
train5000.jsonl |
69 MB | TRAINING_REPORT.json data_sha256 = e6bd9118…1f03dd7 |
eval_development200.jsonl |
4.7 MB | EVALUATION_DEVELOPMENT200.json data_sha256 = febbb313…343ea4b |
eval_reserved158.jsonl |
1.8 MB | EVALUATION_RESERVED158.json data_sha256 = d3f0a72b…1b6be8 |
audit.json |
16 KB | builder's own audit record (passed: true) |
independent-audit.json |
1 KB | second implementation; pins train+eval hashes |
build-provenance.json |
1 KB | script hashes used to build the corpus |
source-reference-silver.jsonl |
4.1 MB | per-section provenance for all 2,208 silver examples |
Training data composition
5,000 eCFR sections (2026-09 snapshot), complexity-balanced:
| Sections | Notes | |
|---|---|---|
| Gold (GPO USLM-derived) | 2,792 | native hierarchy targets |
| Silver (deterministic conversion) | 2,208 | mechanical mapping, not human-reviewed |
| Table-bearing sections | 15 gold / 82 silver |
Silver is a deterministic conversion, not human-reviewed gold — see the model card for the known limitations this introduces.
Evaluation benchmarks
- Development200 — used during development and model selection. Section-stratified across complexity bins.
- Reserved158 — held out until release; never used for any decision. Scored once, published once.
Both files carry input and target per section, in the same message format
as training rows, so the deterministic scorer in the model repo's evaluation
protocol can be re-run directly.
Reproducing the reported numbers
- Download the model repo (weights +
PROMPT_TRAINED.txt). - Generate on
eval_development200.jsonlandeval_reserved158.jsonlwith the trained prompt, temperature and decode settings fromTRAINING_REPORT.json. - Score with the deterministic structure scorer; compare against
EVALUATION_DEVELOPMENT200.jsonandEVALUATION_RESERVED158.jsonin the model repo.
Provenance
All section text derives from public-domain eCFR content (U.S. government work). Gold targets derive from GPO-published USLM XML; silver targets are deterministic conversions of eCFR structure API responses.
Loading and exact bytes
This dataset is public. Native JSONL and source/provenance files retain their original bytes.
The dataset viewer uses stable id / row_json columns for the heterogeneous
span-tree objects; json.loads(row["row_json"]) restores the original record.
Use the original named JSONL files for byte-hash replication. Viewer split roles
match the training/development/evaluation files listed above.
- Downloads last month
- 41