Datasets:

karmendra8386
/

cost-per-correct

task_id stringlengths 2 5	kind stringclasses 9 values	band stringclasses 3 values	variant_of stringclasses 12 values	model stringclasses 2 values	repeat int64 1 1	passed bool 2 classes	prompt_tokens float64 47 186 ⌀	completion_tokens float64 2 512 ⌀	latency_s float64 2.69 567 ⌀	finish_reason stringclasses 3 values
e1	arithmetic	easy	null	qwen2.5-3b-instruct-q4_k_m	1	false	52	5	3.975	stop
e2	extraction	easy	null	qwen2.5-3b-instruct-q4_k_m	1	true	60	5	3.538	stop
e3	classification	easy	null	qwen2.5-3b-instruct-q4_k_m	1	true	60	2	2.69	stop
e4	arithmetic	easy	null	qwen2.5-3b-instruct-q4_k_m	1	true	57	6	3.504	stop
e5	extraction	easy	null	qwen2.5-3b-instruct-q4_k_m	1	true	64	4	3.462	stop
e6	classification	easy	null	qwen2.5-3b-instruct-q4_k_m	1	true	57	2	3.23	stop
e7	structured	easy	null	qwen2.5-3b-instruct-q4_k_m	1	false	61	22	12.101	stop
e8	arithmetic	easy	null	qwen2.5-3b-instruct-q4_k_m	1	true	47	4	2.994	stop
ne1	arithmetic	easy	null	qwen2.5-3b-instruct-q4_k_m	1	false	53	4	3.003	stop
ne2	arithmetic	easy	null	qwen2.5-3b-instruct-q4_k_m	1	true	59	3	3.586	stop
ne3	extraction	easy	null	qwen2.5-3b-instruct-q4_k_m	1	true	72	2	3.903	stop
ne4	classification	easy	null	qwen2.5-3b-instruct-q4_k_m	1	true	62	2	3.006	stop
ne5	structured	easy	null	qwen2.5-3b-instruct-q4_k_m	1	true	74	24	7.665	stop
ne6	arithmetic	easy	null	qwen2.5-3b-instruct-q4_k_m	1	false	65	3	2.803	stop
ne7	classification	easy	null	qwen2.5-3b-instruct-q4_k_m	1	true	65	2	2.798	stop
ne8	extraction	easy	null	qwen2.5-3b-instruct-q4_k_m	1	true	69	3	3.594	stop
ne9	ordering	easy	null	qwen2.5-3b-instruct-q4_k_m	1	true	58	10	5.515	stop
ne10	data	easy	null	qwen2.5-3b-instruct-q4_k_m	1	true	69	3	3.38	stop
m1	analysis	medium	null	qwen2.5-3b-instruct-q4_k_m	1	true	110	48	20.033	stop
m2	analysis	medium	null	qwen2.5-3b-instruct-q4_k_m	1	true	89	25	11.693	stop
m3	analysis	medium	null	qwen2.5-3b-instruct-q4_k_m	1	true	103	301	101.369	stop
m4	multihop	medium	null	qwen2.5-3b-instruct-q4_k_m	1	true	77	82	26.593	stop
m5	structured	medium	null	qwen2.5-3b-instruct-q4_k_m	1	true	68	189	55.488	stop
m6	analysis	medium	null	qwen2.5-3b-instruct-q4_k_m	1	true	111	242	78.389	stop
m7	arithmetic	medium	null	qwen2.5-3b-instruct-q4_k_m	1	false	186	38	25.818	stop
m8	data	medium	null	qwen2.5-3b-instruct-q4_k_m	1	false	100	30	12.592	stop
nm1	arithmetic	medium	null	qwen2.5-3b-instruct-q4_k_m	1	true	61	4	3.1	stop
nm2	data	medium	null	qwen2.5-3b-instruct-q4_k_m	1	true	92	5	5.573	stop
nm3	arithmetic	medium	null	qwen2.5-3b-instruct-q4_k_m	1	false	79	4	4.627	stop
nm4	extraction	medium	null	qwen2.5-3b-instruct-q4_k_m	1	true	78	2	3.488	stop
nm5	structured	medium	null	qwen2.5-3b-instruct-q4_k_m	1	true	65	42	14.372	stop
nm6	classification	medium	null	qwen2.5-3b-instruct-q4_k_m	1	true	108	3	5.123	stop
nm7	data	medium	null	qwen2.5-3b-instruct-q4_k_m	1	false	69	5	5.186	stop
nm8	multihop	medium	null	qwen2.5-3b-instruct-q4_k_m	1	false	72	3	3.096	stop
nm9	data	medium	null	qwen2.5-3b-instruct-q4_k_m	1	true	96	5	5.02	stop
nm10	logic	medium	null	qwen2.5-3b-instruct-q4_k_m	1	false	82	10	5.091	stop
h1	multihop	hard	null	qwen2.5-3b-instruct-q4_k_m	1	true	128	474	191.96	stop
h2	arithmetic	hard	null	qwen2.5-3b-instruct-q4_k_m	1	false	96	2	4.938	stop
h3	multihop	hard	null	qwen2.5-3b-instruct-q4_k_m	1	false	81	512	197.891	length
h4	analysis	hard	null	qwen2.5-3b-instruct-q4_k_m	1	false	137	348	170.433	stop
h5	arithmetic	hard	null	qwen2.5-3b-instruct-q4_k_m	1	false	100	7	6.757	stop
h6	multihop	hard	null	qwen2.5-3b-instruct-q4_k_m	1	false	114	273	100.809	stop
h7	arithmetic	hard	null	qwen2.5-3b-instruct-q4_k_m	1	true	88	372	123.756	stop
h8	multihop	hard	null	qwen2.5-3b-instruct-q4_k_m	1	true	97	512	177.95	length
nh1	analysis	hard	null	qwen2.5-3b-instruct-q4_k_m	1	false	80	5	4.554	stop
nh2	arithmetic	hard	null	qwen2.5-3b-instruct-q4_k_m	1	false	78	6	3.818	stop
nh3	analysis	hard	null	qwen2.5-3b-instruct-q4_k_m	1	false	80	6	4.2	stop
nh4	data	hard	null	qwen2.5-3b-instruct-q4_k_m	1	false	82	398	147.273	stop
nh5	logic	hard	null	qwen2.5-3b-instruct-q4_k_m	1	false	94	228	74.92	stop
nh6	data	hard	null	qwen2.5-3b-instruct-q4_k_m	1	false	81	3	3.871	stop
nh7	analysis	hard	null	qwen2.5-3b-instruct-q4_k_m	1	false	83	5	14.624	stop
nh8	arithmetic	hard	null	qwen2.5-3b-instruct-q4_k_m	1	true	66	5	8.598	stop
nh9	structured	hard	null	qwen2.5-3b-instruct-q4_k_m	1	true	83	22	17.572	stop
nh10	classification	hard	null	qwen2.5-3b-instruct-q4_k_m	1	true	148	3	8.596	stop
p-e1	arithmetic	easy	e1	qwen2.5-3b-instruct-q4_k_m	1	false	50	5	8.029	stop
p-e4	arithmetic	easy	e4	qwen2.5-3b-instruct-q4_k_m	1	false	58	5	7.054	stop
p-e7	structured	easy	e7	qwen2.5-3b-instruct-q4_k_m	1	true	66	24	14.322	stop
p-m6	analysis	medium	m6	qwen2.5-3b-instruct-q4_k_m	1	true	95	399	133.106	stop
p-h6	multihop	hard	h6	qwen2.5-3b-instruct-q4_k_m	1	true	88	134	52.516	stop
p-ne2	arithmetic	easy	ne2	qwen2.5-3b-instruct-q4_k_m	1	false	62	3	16.097	stop
p-ne6	arithmetic	easy	ne6	qwen2.5-3b-instruct-q4_k_m	1	false	59	3	10.262	stop
p-nm3	arithmetic	medium	nm3	qwen2.5-3b-instruct-q4_k_m	1	false	81	3	9.416	stop
p-nm7	data	medium	nm7	qwen2.5-3b-instruct-q4_k_m	1	false	71	3	8.686	stop
p-nm8	multihop	medium	nm8	qwen2.5-3b-instruct-q4_k_m	1	false	65	3	7.232	stop
p-nh1	analysis	hard	nh1	qwen2.5-3b-instruct-q4_k_m	1	false	77	5	12.204	stop
p-nh2	arithmetic	hard	nh2	qwen2.5-3b-instruct-q4_k_m	1	false	73	6	12.987	stop
e1	arithmetic	easy	null	llama-3.2-3b-instruct-q4_k_m	1	false	54	3	10.25	stop
e2	extraction	easy	null	llama-3.2-3b-instruct-q4_k_m	1	true	64	3	8.903	stop
e3	classification	easy	null	llama-3.2-3b-instruct-q4_k_m	1	true	66	2	6.194	stop
e4	arithmetic	easy	null	llama-3.2-3b-instruct-q4_k_m	1	false	61	3	11.508	stop
e5	extraction	easy	null	llama-3.2-3b-instruct-q4_k_m	1	true	66	2	6.005	stop
e6	classification	easy	null	llama-3.2-3b-instruct-q4_k_m	1	true	63	4	7.167	stop
e7	structured	easy	null	llama-3.2-3b-instruct-q4_k_m	1	false	67	14	14.952	stop
e8	arithmetic	easy	null	llama-3.2-3b-instruct-q4_k_m	1	false	50	4	7.783	stop
ne1	arithmetic	easy	null	llama-3.2-3b-instruct-q4_k_m	1	false	53	2	4.505	stop
ne2	arithmetic	easy	null	llama-3.2-3b-instruct-q4_k_m	1	true	63	3	9.773	stop
ne3	extraction	easy	null	llama-3.2-3b-instruct-q4_k_m	1	true	75	2	7.183	stop
ne4	classification	easy	null	llama-3.2-3b-instruct-q4_k_m	1	true	68	2	6.027	stop
ne5	structured	easy	null	llama-3.2-3b-instruct-q4_k_m	1	true	79	15	20.6	stop
ne6	arithmetic	easy	null	llama-3.2-3b-instruct-q4_k_m	1	false	70	2	7.567	stop
ne7	classification	easy	null	llama-3.2-3b-instruct-q4_k_m	1	true	71	3	9.402	stop
ne8	extraction	easy	null	llama-3.2-3b-instruct-q4_k_m	1	false	71	21	26.621	stop
ne9	ordering	easy	null	llama-3.2-3b-instruct-q4_k_m	1	true	62	8	9.688	stop
ne10	data	easy	null	llama-3.2-3b-instruct-q4_k_m	1	false	71	2	7.742	stop
m1	analysis	medium	null	llama-3.2-3b-instruct-q4_k_m	1	true	109	90	81.868	stop
m2	analysis	medium	null	llama-3.2-3b-instruct-q4_k_m	1	true	95	91	80.394	stop
m3	analysis	medium	null	llama-3.2-3b-instruct-q4_k_m	1	true	108	207	182.464	stop
m4	multihop	medium	null	llama-3.2-3b-instruct-q4_k_m	1	true	83	139	140.41	stop
m5	structured	medium	null	llama-3.2-3b-instruct-q4_k_m	1	true	74	454	436.948	stop
m6	analysis	medium	null	llama-3.2-3b-instruct-q4_k_m	1	true	117	198	222.826	stop
m7	arithmetic	medium	null	llama-3.2-3b-instruct-q4_k_m	1	false	165	97	122.475	stop
m8	data	medium	null	llama-3.2-3b-instruct-q4_k_m	1	false	102	134	129.951	stop
nm1	arithmetic	medium	null	llama-3.2-3b-instruct-q4_k_m	1	false	66	2	7.501	stop
nm2	data	medium	null	llama-3.2-3b-instruct-q4_k_m	1	false	86	81	107.776	stop
nm3	arithmetic	medium	null	llama-3.2-3b-instruct-q4_k_m	1	false	82	2	8.852	stop
nm4	extraction	medium	null	llama-3.2-3b-instruct-q4_k_m	1	true	76	2	8.301	stop
nm5	structured	medium	null	llama-3.2-3b-instruct-q4_k_m	1	true	71	67	74.073	stop
nm6	classification	medium	null	llama-3.2-3b-instruct-q4_k_m	1	true	114	3	13.868	stop
nm7	data	medium	null	llama-3.2-3b-instruct-q4_k_m	1	false	73	2	7.009	stop
nm8	multihop	medium	null	llama-3.2-3b-instruct-q4_k_m	1	false	77	2	8.526	stop

End of preview. Expand in Data Studio

Cost Per Correct — v2

A benchmark for evaluating LLM agents on cost-effectiveness rather than accuracy alone. The headline metric is the cost of each correct answer:

cost per correct = total spend (USD) ÷ number of correct answers

A cheap model that is usually right can therefore beat an expensive model that is always right — the trade-off production agent deployments actually face, and the one accuracy-only leaderboards hide. Read the metric together with accuracy: without an accuracy floor it is misleading.

Version 2 (this release, 2026-11-02) replaces the v1 smoke test (5 easy tasks, 17 hosted results) with the expanded, verified task set and the first fully documented pilot results. The Zenodo archival record for this benchmark is DOI 10.5281/zenodo.23439614 (concept DOI 10.5281/zenodo.23439613), deposited 2026-10-17 for v1.

Origin

Created by Karmendra Pandey (ORCID: 0009-0000-3461-9889).
First published 11 October 2026 as a public benchmark on Kaggle Benchmarks: https://www.kaggle.com/benchmarks/karmendra8386/cost-per-correct

The task set (config: `tasks`)

66 entries: 54 independent tasks + 12 paraphrase variants.

Independent tasks: 18 easy / 18 medium / 18 hard. With variants, the full set is 23 easy / 22 medium / 21 hard.
9 kinds: arithmetic (20), analysis (10), data (8), multihop (8), classification (6), structured (6), extraction (5), logic (2), ordering (1).
Each task is a single-turn prompt with a deterministic gold answer and a deterministic, rule-based verifier — no LLM judge anywhere. Golds were established by programmatic recomputation and a 139-check self-test (canonical answers for all 66 tasks, 26 adversarial must-fail probes, independent gold re-derivations).
Paraphrase variants are flagged in variant_of, reword a source task without changing its data or gold, and are never counted as independent tasks.

Files: tasks/tasks_v2.csv and tasks/tasks_v2.json (same 66 tasks; fields: id, kind, band, prompt, expected, variant_of, verifier).

Pilot results (config: `pilot_results`) — exactly what was run

A free, local pilot run on 2026-11-02 — zero spend, no API keys:

Models (both local, Q4_K_M GGUF, run on llama.cpp llama-server): Qwen2.5-3B-Instruct and Llama-3.2-3B-Instruct. These are small local open-weight models — the pilot results are not results for any hosted or frontier model.
Qwen2.5-3B-Instruct: all 66 tasks — 35/66 passed (53.0%); independent tasks 32/54 (59.3%); 5,311 prompt + 4,918 completion tokens.
Llama-3.2-3B-Instruct: a 41-task subset — 22/41 passed (53.7%); 3,276 prompt + 2,878 completion tokens. The Llama run covers all 9 kinds and all 3 bands; it was stopped at 41 tasks by the operator (host CPU contention), so tasks h6–h8, nh1–nh10 and all 12 variants have no Llama result. One Llama row (h4) is a recorded error (client timeout, no graded answer).
Settings: temperature 0, max_tokens 512, 1 repeat per task.
Token counts are the runtime's own reported usage — measured, not estimated. Latency is client-observed wall time on a CPU-contended shared VM and should not be compared across models.
No API dollar costs are claimed for these runs: actual cost was $0.00 (local inference). Where a hosted-equivalent cost per correct is quoted in the summary (≈ $0.000019 Qwen / $0.000018 Llama), it prices the measured tokens at a dated third-party list price (Groq llama-3.1-8b-instant, $0.05 in / $0.08 out per 1M tokens, snapshot 2026-10-31) and is labelled an estimate.
Proof: correctness was re-graded from the saved raw responses with the dataset's own verifiers — 107 of 107 verdicts reproduced. See proof/PROOF.md (model file SHA256s, runtime, timeline, full disclosure of incidents), proof/summary.md, proof/grading_log.csv, and results/raw_outputs.jsonl (full prompt + full raw response for every run). results/pilot_results.csv is the full results table including response text; the viewer config uses results/pilot_results_viewer.csv, the same 107 rows without the response-text column.

Limitations

66 tasks is still a small benchmark; read results as cost-efficiency signals, not general model rankings.
The pilot covers two small local 3B models only, one run per task, and Llama covers only 41 of the 66 tasks.
Hosted-equivalent dollar figures depend on a dated price snapshot and are comparable only across runs priced at the same rates.

Citation

@misc{pandey2026costpercorrect,
  author       = {Pandey, Karmendra},
  title        = {{Cost Per Correct}: Benchmark Tasks for
                  Cost-per-Correct-Answer Evaluation of {LLM} Agents},
  year         = {2026},
  version      = {2.0.0},
  howpublished = {Kaggle Benchmarks / Hugging Face Datasets},
  url          = {https://www.kaggle.com/benchmarks/karmendra8386/cost-per-correct},
  doi          = {10.5281/zenodo.23439614},
  note         = {Zenodo dataset record, deposited 2026-10-17 (v1); v2 task set and local pilot results published on Hugging Face 2026-11-02}
}

Downloads last month: 109

Cost Per Correct — v2

Origin

The task set (config: tasks)

Pilot results (config: pilot_results) — exactly what was run

Limitations

Citation

Space using karmendra8386/cost-per-correct 1

The task set (config: `tasks`)

Pilot results (config: `pilot_results`) — exactly what was run