superlinked/train-a-router-bbeh

Train a router on Big-Bench Extra Hard

This release accompanies the self-contained SIE train-a-router example. It contains the questions, observed answer-model outcomes, fixed embeddings, trained decision models and measurements needed to reproduce its cost–quality comparisons.

On 975 held-out questions, an embedding classifier with learned cost estimates saved 16.7% with a 0.9-point accuracy loss; another preselected policy saved 39.7% with a 6.1-point loss. Both savings figures include the Cloud encoder fee. These are held-out results from policies frozen before final answers were collected. The full curves are supplied so users can choose their own spending/accuracy tradeoff.

Data and labels

The source is Google DeepMind's Big-Bench Extra Hard, revision 80d12ca916b7158f22293fcf3144f4d3d854d4be. Source-only duplicate, family and conflicting-reference checks retain 3,918 questions from 23 tasks. Identified source/context families never cross splits.

Split	Questions	Use
train	1,969	Fit correctness and cost predictors
development	589	Select model settings and training duration
calibration	385	Choose routing policies
test	975	Evaluate frozen policies

Both gpt-6-luna and gpt-6.1-sol answered every retained question independently with medium reasoning, a 16,384-output-token limit and the same exact-answer format. The first terminal response is retained; truncated answers count as incorrect. Correctness comes from checking actual generated answers against BBEH references, not from a teacher guessing a route. Source-validated answer-format normalization and the original official grading result are preserved.

Each JSONL row has id, task, family, question, reference_answer, luna_correct, sol_correct, luna_usd and sol_usd. The richer artifacts bind model identities, selected responses, usage, source revision, grading code, feature hashes and training choices. Their immutable SHA-256 values are listed in manifest.json.

# Load the public question/outcome tables with the Hugging Face datasets library.
from datasets import load_dataset
data = load_dataset("superlinked/train-a-router-bbeh")

Use the example's pinned release.json and router.py fetch for the exact published experiment rather than following the moving dataset branch.

Compared recipes

• Unchanged Laya: a roughly 421-million-parameter ModernBERT decision model, queried separately about Luna and Sol correctness.
• Laya with LoRA: rank-16 attention updates and the existing decision head. Development selected three training passes; the model was then restarted and fitted on train + development.
• Embedding neighbors: average observed outcomes of 32 similar training questions.
• Embedding classifiers: two regularized logistic regressions on fixed 2,560-dimensional Qwen3-Embedding-4B vectors.
• Cost variants: each quality predictor is also evaluated with shared text/length predictors of per-model answer cost.
• Controls: text/length, text-predicted task, true task metadata, always-Luna, always-Sol and random routing. True task metadata is explicitly marked as metadata-assisted.

Laya is pinned to convaiinnovations/laya revision aa8c91ca088ec597df95a0d1c76b3063cb2ae5e8. Qwen is pinned to Qwen/Qwen3-Embedding-4B revision 5cf2132abc99cad020ac570b19d031efec650f2b. The Qwen encoder stays fixed; its weights are not included here. Original Laya and merged adapted weights are included.

Results and interpretation

Method	Answer accuracy	Accuracy loss	Answer cost / 10k	Answer-cost savings
Always Sol	80.8%	0.0 points	$70.77	0.0%
Always Luna	55.6%	25.2 points	$12.23	82.7%
Unchanged Laya	68.6%	12.2 points	$41.82	40.9%
Laya + LoRA	78.3%	2.6 points	$56.80	19.7%
Embedding neighbors	76.3%	4.5 points	$53.80	24.0%
Embedding classifier	76.8%	4.0 points	$54.41	23.1%
LoRA + learned cost	79.0%	1.8 points	$57.63	18.6%
Embedding classifier + learned cost	79.9%	0.9 points	$58.16	17.8%
Text/length + learned cost	79.0%	1.8 points	$55.28	21.9%
Text-predicted task + learned cost	80.1%	0.7 points	$58.31	17.6%

All router rows use a rule fixed before final collection: select the highest calibration accuracy under an 80% calibration answer-cost cap relative to always-Sol. It is an illustrative spending level, not a required quality threshold or a guarantee of 20% test savings. The table reports final-test answer costs; add $0.81 per 10,000 requests for the Cloud encoder on embedding routes. For embedding + learned cost, that changes $58.16 to $58.97, or 16.7% savings versus always-Sol at $70.77. Native and application hosting are separate. Every method, both cost variants and all preselected spending levels remain in the full report.

Answer costs use measured selected-response token usage and the price schedule pinned on 2026-10-21. The report separately shows projected SIE Cloud singleton encoder fees validated against 245 batches and 91 singleton calls; observed acquisition charges are retained separately. Native serving and application hosting are not treated as free; measured latency and an explicitly separate busy-resource hosting scenario are included. One-time acquisition/training costs are separate from per-request serving cost.

The final report includes all saved calibration candidates, 2,000 family-within-task bootstrap draws for pointwise uncertainty intervals, and random-routing comparisons. No final-test threshold or model selection is performed. A source-format sensitivity check removes two questions flagged before final collection and replays the same policies on 973 questions.

This tests new questions within the observed 23-task mixture, not unseen reasoning tasks. Public benchmark contamination in answer-model pretraining is not ruled out. Test labels are now public: a replay or refit using this release is a reproduction, not a new blind evaluation. A teacher-labeled deployment needs independently checked evaluation labels; teacher accuracy was not established by this experiment.

Artifact groups

• core: questions, references, outcome labels, embeddings, small fitted heads, routing policies, reports and provenance.
• weights: the original and adapted Laya model trees.
• recordings: sanitized paired answer collection records, including selected response usage and retained transport-attempt accounting.

All runnable code lives in the example folder; the dataset stores large evidence and trained artifacts. The one-page instructions give the execution sequence.

Attribution and licenses

BBEH's source software is Apache-2.0; its other materials are CC-BY-4.0. This dataset preserves the upstream README, citations and license under source/provenance/. Please credit the BBEH authors and consult that README for the paper citation and task-specific attribution. The tabular data and derived question/outcome artifacts are distributed with that attribution under CC-BY-4.0.

Laya and Qwen model cards declare Apache-2.0. The original Laya model tree and its license notices are retained, and the LoRA adaptation is identified as a modification. Model cards, exact revisions and a component-level attribution record are under metadata/licenses/ and metadata/attribution.json. The Laya weights retain their Apache-2.0 license; the dataset-card license does not replace component licenses.

Answer recordings contain model-generated outputs, not human-written reference answers. Credentials, environment dumps and provider-encrypted reasoning are excluded. This release contains public benchmark material and synthetic model responses; it contains no user production queries.