These artefacts are released for evaluation and research. They are distilled from commercial model outputs and built on third-party corpora; you are responsible for checking that your use complies with those upstream terms.

GraceX query-expansion evaluation set

The 1,000 held-out inputs every number in this project's model cards is measured on, across 39 languages.

Part	Items	Where it comes from
`synthetic_heldout`	500	the test split of the generated data, held out by subtopic so a paraphrase of a training query cannot appear here
`real_unseen`	500	real user queries from the public `fineinstructions/real_queries` corpus (WildChat), filtered to those the production service would expand, and checked against the training data

Checksum of the set: 22a12a65bd0d6950.

Fields

{"id": "syn-q006051", "source": "synthetic_heldout", "language": "zh",
 "payload": {"query": "...", "contextDomains": [], "conversationContext": [...]},
 "prompt": "<the production prompt, rendered from the payload>",
 "reference": "<the held-out reference answer, synthetic part only>",
 "template": "statement", "domain": "arts & crafts",
 "has_history": true, "has_context_domains": false, "needs_coref": true}

payload is exactly what the production service passes to its expansion call, and prompt is that payload rendered through the production prompt builder, so a system under test sees the same text production would send it. reference is the held-out target for the synthetic part; the real queries have no reference, which is why comparisons are judged pairwise rather than scored against a gold answer.

How it is used

Each system answers every input with the production prompt at temperature 0.3. Two kinds of measurement follow:

The output contract, checked exactly — a single line, the input reproduced verbatim, something appended, no preamble or markdown, same language as the input — with the same validator that accepted the training targets, plus statistical language identification. Inputs that cannot satisfy the contract (those containing a newline) are reported separately.
Judged comparisons — an independent LLM judge from neither the teacher's nor the reference model's family sees two answers, in both presentation orders, and may call a tie; results are reported as a net win rate with a bootstrap confidence interval and a sign test.

The harness that does this (eval/qe_eval.py) is part of the source repository. No stub systems and no keyword heuristics are used anywhere in it.

Caveat

The set is majority English, so results on it describe the case these models handle best. A deployment whose traffic is mostly non-English should re-weight it by the real language mix before any decision.

Licensing and provenance — read before use

This is a derivative work with several upstream constraints, and none of them has been legally reviewed:

Base weights: Qwen/Qwen3.5-0.8B. The base model's licence governs the weights; check it before redistribution or commercial use.
Training targets are distilled from commercial models accessed through OpenRouter (a GLM model for the core data, gemini-3.1-flash-lite for the additional batches). Provider terms may restrict using such outputs to train models, and those terms are the user's responsibility.
Source corpora: EmbeddingStudio synthetic search queries (Apache-2.0), MIRACL queries (Wikipedia-derived), BeIR/CQADupStack generated queries (StackExchange-derived), and a general instruction set used as replay data. Each carries its own licence.
Evaluation inputs include real user queries from a public corpus, filtered to those the production service would expand.

The repository is gated for this reason: access is granted per request, and the recipient accepts responsibility for upstream compliance.

Downloads last month: -

Total file size:

2.87 MB