Datasets:
license: other
license_name: polyform-noncommercial-1.0.0
license_link: LICENSE
language:
- it
- en
task_categories:
- text-generation
pretty_name: SemanticRepair bilingual native routing challenge (2026-10-05)
size_categories:
- n<1K
tags:
- evaluation
- query-rewriting
- routing
- negation
- multi-intent
configs:
- config_name: default
data_files:
- split: test
path: data/test.jsonl
SemanticRepair bilingual native routing challenge
Reproduction warning: the historical runner in source/ creates hardlinks to model artifacts. Use the corrected independent-copy runner in compound/source/bench/semantic_repair/run.py; the incident and cleanup evidence are documented in compound/README.md. These scores are historical and the banks are development/diagnostic material after output inspection.
160 newly authored questions, 80 paired Italian/English anchors, with frozen desired capability sets and separate request-presence labels. The bank measures the Q8_0 SemanticRepair-270M v22 model and its admission through SDG's general default graph. It is a self-evaluation authored by Codex / Atlante for Gramscii, without independent label review.
Results, 5 October 2026
| Measure | Native direct pre-repair gate | Native Chat with repair |
|---|---|---|
| Complete desired set or correct refusal, all 160 | 21.875% (35/160) | 33.750% (54/160) |
| Exact supported requests, 134 | 19/134 | 47/134 |
| Correct graph refusals, 26 | 16/26 | 7/26 |
| Exact multi-capability requests, 40 | 0/40 | 1/40 |
| Admissions containing an unwanted capability | 24/160 | 54/160 |
| Median routing latency, startup excluded | 32.4 ms | 174.8 ms |
Repair increases exact request coverage, while unwanted admissions increase from 24 to 54. It is not a safety improvement on this bank. A wrong admission is a routing decision: no tool executes, so it is not evidence of an actual download, deletion, email, or other action. Strict subsets of a requested set remain partial, never exact.
| Family (20 items each) | Direct gate correct | Native Chat correct |
|---|---|---|
| injection | 20.0% | 40.0% |
| multi_three | 0.0% | 0.0% |
| multi_two | 0.0% | 5.0% |
| negation | 30.0% | 70.0% |
| noisy | 15.0% | 40.0% |
| plain | 35.0% | 65.0% |
| quoted_trap | 15.0% | 30.0% |
| uncovered | 60.0% | 20.0% |
Native Chat correctness is 37.5% in Italian and 30.0% in English, using their actual separate workspace bars. The English bars are not recalibrated for this bank. This comparison includes native admission rules and cannot isolate the causal contribution of the repair model alone.
Model-only measurements
Raw completion request recall: 131/150 (87.33%). Correct nonrequest detection: 9/10 (90.0%). All annotated protected strings appear in 59/72 (81.94%) of the applicable normalized completions. Raw output has the desired number of request lines in 33/40 multi-capability items (17/20 two-request, 16/20 three-request). Line count does not establish semantic decomposition quality.
Unsupported actions are still requests for this raw intent-presence metric. They must be refused by graph admission, not mislabeled as nonrequests. Literal protected-string retention is stricter than semantic equivalence; it does not measure argument retention by executing tools, which consume original input.
Scope and reproducibility
The SDG base is 2d01f906 (including #645 and #654), with the benchmark source at 34996e765a0b13f5dae85efbd82ef98aa62c2efa. Model weights are unchanged: sft-v22-q8_0.gguf, SHA-256 58b750b4bba4ed9cf1ea30316f1c5ad131316efaedd56cfd9440b488b2ae88f4, model revision acc30fd451d58620cc22dbbfab46caed3cbf12f5. Inference uses the pinned llama.cpp build 10621 on macOS arm64 Metal, greedy completion, 96-token maximum, repeat penalty 1.0, and the native normalization policy. The embedder is EmbeddingGemma-300m BF16; all artifact pins, graph descriptions, bars, platform, and evidence are in provenance.json and graphs.json.
Every item gets a fresh conversation with no selected source documents and web enabled. The source database starts empty, including native empty OpenData search stores. Web providers, graph tools, narration, tender downloads, Regolo generation, and source processing do not execute. Database backups remain private. The isolated model processes are stopped and collected; their listeners are released, and the exact test database is backed up and dropped.
Reproduce using the tested SDG source revision, the published workspace configuration, and the runner instructions. Select graph default explicitly. provenance.json carries the plugin switches and host declarations; host declarations name permission and credential environment variables, never secret values. Start from an empty migrated database. The runner configuration schema requires explicit local paths and ownership. No published source documents or private workspace content are needed for this admission-only experiment.
challenge.json is the byte-identical frozen bank, SHA-256 5274284481d2f8544c732da024befa4710908fecd35ce14c685228953dcad4b9. data/test.jsonl exposes the same 160 rows to dataset tools. manifest.json records the overlap audit: no exact normalized question overlap with v22 train/validation/test, prior probes, or the older 481-question routing bank. Exact exclusion does not prove semantic independence. Gold labels and thresholds are not changed after observing outputs; the model is not retrained.
results.json contains every aggregate; traces.jsonl contains original questions, annotations, direct routes, native decisions, refusals, normalized raw completions, and latency. Source scorer is included. The focused gate passes 9 scoring tests, 1 real native integration test processing all 160 rows, and Ruff. The product's full suite is not rerun because this change adds benchmark files and leaves product code and dependencies unchanged.
Limitations
Eight equally sized families are deliberately adversarial, not representative production frequencies. The 80 translated pairs are dependent observations. Published Wilson intervals and the item-level sign-test p-value assume independence; they must not be promoted as independent population inference. No independent annotation review, repeated-run stability study, populated-catalogue evaluation, source-grounded answer evaluation, tool execution, or conversation-continuation benchmark is provided. Only Italian and English are tested here.
This dataset is an ordinary published evaluation dataset. It does not claim a Hugging Face verified-evaluation badge or registration in the Hub's beta benchmark leaderboard service.
License and attribution
Original challenge content, contracts, and evaluator software retain the repository's PolyForm Noncommercial 1.0.0 terms and Gramscii copyright notice in LICENSE. Model weights remain under Gemma terms and are not distributed in this dataset. Questions refer to hypothetical identifiers and public place names; no private documents or production conversations are included.