nl2jq-bench

Part of the nl2jq project — all artifacts · code + CLI · live demo · dataset

An execution-scored, frozen benchmark for natural-language → jq program generation. Each item is a realistic request over a small JSON input with a verified reference program; models are scored on whether their program's output matches the reference output under jq 1.7.1 — not on string similarity.

Files

file	what
nl2jq-bench-1.0.0.jsonl	the benchmark: 400 frozen items (v1.0.0)
FREEZE.txt	the freeze record — content sha256 (the benchmark's identity), tier histogram, jq binary hash
harness.py	standalone scorer (Python 3.10+, any jq binary; pin jq 1.7.1 to reproduce)
devset-v0.jsonl	the RETIRED 100-item development split — see "History" below; not a benchmark

Cite scores as nl2jq-bench@1.0.0 together with the sha256 in FREEZE.txt. Any change to the items is a new version with a new hash; v1.0.0 never changes.

Design

400 items, five tiers (each item carries tier = difficulty 1–5):

tier	n	content
T1 Elementary	60	single op: path/index/slice/length/keys/has/type///
T2 Core	100	select/map, comparisons, arithmetic, unique/sort/min/max/add, contains
T3 Aggregation & reshape	100	group_by/sort_by/*_by, to_entries family, map_values, flatten, any/all, object construction
T4 Strings, formats & composition	80	split/join/test/capture/sub/gsub, casts, interpolation, @csv/@tsv/@json/@base64, 3+-stage pipelines
T5 Generalization	60	reduce, foreach, walk/.., paths/getpath, try/catch, if/then/elif, object-merge add, computed-key group_by, INDEX

Anti-contamination, by construction (relative to the nl2jq training set):

• Field-name disjointness: 0% overlap. Every input field name was checked against the full training vocabulary (687 names incl. constructible compounds) with a near-miss bar (no shared ≥4-char prefixes, no small-edit variants, no common abbreviations) and an enum-value gate (no training enum value appears as an input value or program literal). The gate ships in the project repo (bench/validate_novelty.py) and rebuilds the forbidden vocabulary from the generator source at check time.
• T5 is beyond-grammar. The 60 T5 items use only constructs the training-data grammar provably never emits (verified by an emission audit + a source gate, bench/audit_coverage.py). Report the T5-only score separately — it is the generalization headline; a model cannot have seen these operation shapes in nl2jq.
• Novel domains. 12 domains disjoint from the training set's (clinical, gradebook, transit, sports, music, weather, real-estate, recipes, library, gaming, lab, civic), plus ≤2% abstract single-letter-key items (tagged abstract: true).
• Sealed canary. A 100-item sealed set was frozen alongside v1.0.0 (sha256 in FREEZE.txt, never published). It exists to detect future train-on-test: a large public−sealed gap for a model is a contamination signal. Every item also carries a unique canary_id for corpus grepping.

Review process. Every item was execution-verified under a pinned jq 1.7.1 binary (expected outputs are pinned from execution, never typed), then adversarially reviewed for ambiguity: where a different but defensible program produces a different output, the request was disambiguated or the alternative's executed output added to acceptable_outputs. 11% of items were revised in this pass.

Item format

{
  "id": "v1-042", "tier": 3, "difficulty": 3, "domain": "transit",
  "request": "total ridership per garage",
  "input": [ {"route_no": "41B", "garage": "Elmside", "ridership": 10432}, ... ],
  "reference_program": "group_by(.garage) | map({garage: .[0].garage, ridership: (map(.ridership) | add)})",
  "expected_output": [ ... ],
  "order_insensitive": false, "abstract": false,
  "novel_fields": ["route_no", "garage", "ridership"],
  "grammar_covered": true, "provenance": "hand", "source": "original",
  "jq_version": "1.7.1", "frozen_in": "1.0.0", "canary_id": "…", "license": "CC-BY-4.0"
}

Some items carry acceptable_outputs (a list of equally-correct output streams).

Scoring

A prediction is correct if its jq output equals an acceptable output under these normalizations (see harness.py):

• array/stream equivalence — map(.x) and .[].x score the same;
• array-order-insensitivity for items flagged order_insensitive;
• float tolerance for numeric outputs.

harness.py is standalone — it needs only Python 3.10+ and a jq binary (uses $JQ_BIN, else jq on PATH; pin jq 1.7.1 to reproduce published numbers exactly):

from harness import score_program
score_program('map(.route_no)', item)   # -> {"valid": True, "correct": ...}

One-shot discipline: this benchmark is meant to be evaluated once per model release. Do not select checkpoints, tune prompts, or iterate data against it — that is what devset-v0.jsonl is for.

Results (v1.0.0)

The complete run-by-run record lives in this repo's RESULTS.md (append-only; one row per official run). Summary as of 2026-07-11 — rows marked (superseded) are earlier fine-tunes kept for the record; the shipped models are the bold rows:

system	pass@1	valid@1	T1	T2	T3	T4	T5-gen
40m v5, from scratch (superseded)	0.00	0.48	0.00	0.00	0.00	0.00	0.00
nl2jq-40m (v7 weights)	0.04	0.56	0.08	0.08	0.02	0.01	0.02
nl2jq-40m + input-grounded decoding	0.09	0.76	0.17	0.14	0.08	0.01	0.05
Qwen3-0.6B base, zero-shot	0.01	0.06	0.02	0.00	0.00	0.00	0.02
0.6B v5 LoRA (superseded)	0.24	0.67	0.45	0.38	0.21	0.12	0.02
nl2jq-qwen3-0.6b (v6 weights)	0.40	0.73	0.65	0.52	0.36	0.28	0.20
nl2jq-qwen3.5-2b (v7 LoRA)	0.46	0.75	0.67	0.61	0.47	0.33	0.13
nl2jq-qwen3.5-2b + exec-rerank k=4 (CLI config)	0.48	0.82	0.75	0.61	0.47	0.35	0.18
Claude Opus 4.8 (zero-shot)	0.96	0.98	1.00	0.96	0.96	0.96	0.90

For contrast, on the retired in-distribution dev split the earlier models scored 0.55 / 0.81 / 0.75 pass@1. The differences are the measured generalization gap, and quantifying it honestly is the point of this benchmark. Three regimes emerge:

• From scratch at 37M: in-distribution skill barely transfers (dev 0.55 → frozen 0.00–0.09 across three data generations). Syntax survives the distribution shift; field binding and composition do not.
• Pretrained bases + task LoRA: fine-tuning does all the work (base 0.01 → 0.40 at 0.6B; 0.48 at 2B in the CLI configuration) — strong on elementary tasks over never-seen fields, progressively weaker on multi-stage composition and the T5 constructs.
• Frontier zero-shot climbs from 0.75 to 0.96: a share of its dev-split misses were ambiguity artifacts that this benchmark's adversarial review eliminated (disambiguated requests, executed acceptable_outputs).

History: devset-v0

The original 100-item set (shipped here as devset-v0.jsonl) was used during development for checkpoint selection and grammar iteration, and its field names overlap the training vocabulary (~65%) — a post-hoc audit found 95/100 of its items structurally reproducible by the training-data generator. It is retired as a benchmark and kept only for reproducibility of the development history. Do not report devset numbers as held-out results.

Provenance

All items are original, authored for this benchmark (no scraped content, no Stack Overflow), execution-verified under jq 1.7.1. Reference programs and inputs are CC BY 4.0.

Code: github.com/gauthierpiarrette/nl2jq. Companion training set: gauthierpiarrette/nl2jq. Models: nl2jq-40m, nl2jq-qwen3-0.6b, nl2jq-qwen3.5-2b.