BudgetBench™

BudgetBench tests budgeting chat assistants. A person asks questions about their own bank transactions. The assistant answers in a chat. BudgetBench checks every figure in every answer.

Version 1.0.1, published 2026-10-20 as the public dataset PulseOS/budgetbench. Grader v1.0.2 and oracle v1.0.3 (2026-10-22) are in this repository and not yet published; see CHANGELOG.md.

What it measures, and why

A budgeting assistant has one first duty: it must never show a wrong figure about the person's money. A wrong total that looks right is worse than no answer. Its second duty is to give the right answer, with no tap or after one tap on an option it offers.

Most tests count right answers only. A system that guesses then scores well and still shows wrong figures. Thus BudgetBench gives each case exactly one of six labels and reports all six separately. It never merges them into one score:

label	meaning
Wrong shown	the answer shows a wrong figure, name or claim
Wrong offered	the system asks which reading you mean, and every option it offers is wrong
Right, no tap	every turn is right, with no tap
Right after one tap	every turn is right after one tap on an offered option
Not sure	the system says it is not sure, or asks back
Declined	the system declines or answers something else, with no figure

A seventh state, needs review, is not a result: the grader could not place a figure, and a person reads the row (METHODOLOGY.md §6).

Summary

BudgetBench has 18 synthetic households in the real shapes of bank and app exports (a connected bank, Monarch, YNAB, Chase, Amex, Capital One, Copilot, German bank and card files in euros, a plain CSV, and two-currency households) and 1,060 conversations of 1 to 4 turns about them: totals, shops, months, comparisons, exclusions, follow-up chains, how-to questions about the app, requests the assistant must refuse, and currencies. An independent oracle computes every gold answer from the household's truth; it shares no code with Pulse Local, and only a CSV line reader (households/csv.mjs) with the reference tools of the model rows. A rule-based grader, with no model judge, checks every money figure of every turn. The public split (723 cases on 12 households) is for development. A sealed split (337 cases on 6 households, written by independent writers who never saw any app code) stays private and supports release claims; only its hashes are public.

Leaderboard

Each cell: cases (share, 95% interval from a bootstrap that resamples households, then cases). The public split is a DEVELOPMENT split: 494 of its 723 cases are Pulse Local's earlier test questions, or the same wordings on new households (CONTAMINATION.md). The second table shows the 229 public cases from independent writers: the chat code of Pulse Local 0.4.16 did not change after anyone read its answers to them (while fixing the app after 0.4.16, its makers read some of them, so the 0.4.19 rows are development scores here too), and nobody tuned the stock models. They come from 4 households, so their intervals are wide. The third table is the sealed split: aggregate counts only.

system	tuned on public	n	wrong shown	wrong offered	right, no tap	right after one tap	not sure	declined	needs review	run
Pulse Local 0.4.16, rules	yes (494 exposed cases)	723	39 (5.4%, 2.0% to 9.0%)	3 (0.4%, 0.0% to 1.2%)	604 (83.5%, 74.5% to 93.1%)	2 (0.3%, 0.0% to 0.8%)	55 (7.6%, 3.1% to 12.2%)	19 (2.6%, 0.7% to 4.9%)	1	2026-10-19
Pulse Local 0.4.16, rules + 4B checker (agree or not sure)	yes (494 exposed cases)	723	11 (1.5%, 0.3% to 3.2%)	4 (0.6%, 0.0% to 1.5%)	566 (78.3%, 69.4% to 88.8%)	2 (0.3%, 0.0% to 0.8%)	126 (17.4%, 9.3% to 23.8%)	14 (1.9%, 0.5% to 3.7%)	0	2026-10-19
Pulse Local 0.4.19, rules	yes (494 exposed cases, and more read after 0.4.16)	723	31 (4.3%, 1.7% to 7.1%)	1 (0.1%, 0.0% to 0.7%)	618 (85.5%, 77.4% to 93.7%)	2 (0.3%, 0.0% to 0.9%)	48 (6.6%, 2.8% to 10.5%)	21 (2.9%, 0.7% to 5.5%)	2	2026-10-25-0419, 2026-10-25
Pulse Local 0.4.19, rules + 0.6B checker (agree or not sure)	yes (494 exposed cases, and more read after 0.4.16)	723	5 (0.7%, 0.0% to 1.9%)	2 (0.3%, 0.0% to 1.1%)	560 (77.5%, 69.2% to 86.2%)	2 (0.3%, 0.0% to 0.9%)	138 (19.1%, 12.3% to 25.0%)	15 (2.1%, 0.5% to 4.1%)	1	2026-10-25-0419, 2026-10-25
Qwen3 0.6B (stock, Q4_K_M, thinking off) + reference tools	no	723	199 (27.5%, 22.6% to 32.9%)	0 (0.0%, 0.0% to 0.0%)	55 (7.6%, 5.5% to 10.1%)	0 (0.0%, 0.0% to 0.0%)	1 (0.1%, 0.0% to 0.7%)	457 (63.2%, 57.8% to 68.8%)	11	2026-10-19
Qwen3 4B (stock, Q4_K_M, thinking off) + reference tools	no	723	303 (41.9%, 37.3% to 47.6%)	1 (0.1%, 0.0% to 0.6%)	161 (22.3%, 16.8% to 26.5%)	0 (0.0%, 0.0% to 0.0%)	10 (1.4%, 0.2% to 2.9%)	229 (31.7%, 26.9% to 36.6%)	19	2026-10-19
Qwen3 8B (stock, Q4_K_M, thinking off) + reference tools	no	723	267 (36.9%, 30.2% to 41.9%)	0 (0.0%, 0.0% to 0.0%)	287 (39.7%, 34.2% to 48.1%)	0 (0.0%, 0.0% to 0.0%)	20 (2.8%, 1.0% to 4.4%)	130 (18.0%, 14.2% to 21.9%)	19	2026-10-19

Writers' cases only (the 229 public cases by independent writers; the chat code of Pulse Local 0.4.16 did not change after anyone read its answers to them, but 0.4.19's makers read some of them while fixing the app; nobody tuned the stock models; 4 households, so the intervals are wide):

system	tuned on public	n	wrong shown	wrong offered	right, no tap	right after one tap	not sure	declined	needs review	run
Pulse Local 0.4.16, rules	yes (494 exposed cases)	229	35 (15.3%, 10.2% to 20.4%)	3 (1.3%, 0.0% to 3.1%)	123 (53.7%, 43.2% to 62.9%)	1 (0.4%, 0.0% to 1.7%)	48 (21.0%, 14.4% to 28.8%)	18 (7.9%, 3.9% to 12.8%)	1	2026-10-19
Pulse Local 0.4.16, rules + 4B checker (agree or not sure)	yes (494 exposed cases)	229	10 (4.4%, 1.3% to 8.7%)	4 (1.7%, 0.0% to 3.9%)	108 (47.2%, 33.8% to 59.9%)	1 (0.4%, 0.0% to 1.7%)	93 (40.6%, 30.2% to 52.1%)	13 (5.7%, 2.7% to 9.3%)	0	2026-10-19
Pulse Local 0.4.19, rules	yes (494 exposed cases, and more read after 0.4.16)	229	26 (11.4%, 6.3% to 17.1%)	1 (0.4%, 0.0% to 1.7%)	137 (59.8%, 48.7% to 69.0%)	2 (0.9%, 0.0% to 2.7%)	41 (17.9%, 12.0% to 25.1%)	20 (8.7%, 4.3% to 14.0%)	2	2026-10-25-0419, 2026-10-25
Pulse Local 0.4.19, rules + 0.6B checker (agree or not sure)	yes (494 exposed cases, and more read after 0.4.16)	229	5 (2.2%, 0.0% to 5.5%)	2 (0.9%, 0.0% to 3.0%)	117 (51.1%, 40.7% to 61.3%)	2 (0.9%, 0.0% to 2.7%)	88 (38.4%, 30.6% to 46.8%)	14 (6.1%, 2.7% to 10.5%)	1	2026-10-25-0419, 2026-10-25
Qwen3 0.6B (stock, Q4_K_M, thinking off) + reference tools	no	229	48 (21.0%, 13.8% to 29.7%)	0 (0.0%, 0.0% to 0.0%)	12 (5.2%, 2.2% to 8.9%)	0 (0.0%, 0.0% to 0.0%)	1 (0.4%, 0.0% to 1.8%)	164 (71.6%, 61.5% to 79.9%)	4	2026-10-19
Qwen3 4B (stock, Q4_K_M, thinking off) + reference tools	no	229	88 (38.4%, 30.5% to 47.1%)	1 (0.4%, 0.0% to 2.0%)	42 (18.3%, 11.0% to 24.9%)	0 (0.0%, 0.0% to 0.0%)	9 (3.9%, 0.9% to 7.5%)	78 (34.1%, 27.5% to 41.0%)	11	2026-10-19
Qwen3 8B (stock, Q4_K_M, thinking off) + reference tools	no	229	108 (47.2%, 38.8% to 57.4%)	0 (0.0%, 0.0% to 0.0%)	51 (22.3%, 14.2% to 30.1%)	0 (0.0%, 0.0% to 0.0%)	15 (6.6%, 3.3% to 10.3%)	44 (19.2%, 13.5% to 24.8%)	11	2026-10-19

Sealed split (337 cases, 6 households, held out; aggregate counts only):

system	tuned on public	n	wrong shown	wrong offered	right, no tap	right after one tap	not sure	declined	needs review	run
Pulse Local 0.4.16, rules	yes (494 exposed cases)	337	79 (23.4%, 16.9% to 29.9%)	1 (0.3%, 0.0% to 1.2%)	178 (52.8%, 45.8% to 59.9%)	3 (0.9%, 0.0% to 2.6%)	50 (14.8%, 9.8% to 20.2%)	22 (6.5%, 3.0% to 10.6%)	4	2026-10-19
Pulse Local 0.4.16, rules + 4B checker (agree or not sure)	yes (494 exposed cases)	337	23 (6.8%, 2.5% to 11.0%)	1 (0.3%, 0.0% to 1.2%)	153 (45.4%, 37.9% to 52.7%)	3 (0.9%, 0.0% to 2.6%)	143 (42.4%, 35.3% to 49.6%)	14 (4.2%, 1.8% to 6.9%)	0	2026-10-19

After review of the needs-review rows (rulings by a Claude session for the maintainers, in each run folder's ADJUDICATION.md; the grader's counts above stay the reproducible ones):

system	split	n	wrong shown	wrong offered	right, no tap	right after one tap	not sure	declined	needs review
Pulse Local 0.4.16, rules	public	723	39	3	605	2	55	19	0
Pulse Local 0.4.16, rules + 4B checker (agree or not sure)	public	723	11	4	566	2	126	14	0
Pulse Local 0.4.19, rules	public	723	31	1	619	2	48	22	0
Pulse Local 0.4.19, rules + 0.6B checker (agree or not sure)	public	723	5	2	560	2	138	16	0
Qwen3 0.6B (stock, Q4_K_M, thinking off) + reference tools	public	723	207	0	55	0	1	460	0
Qwen3 4B (stock, Q4_K_M, thinking off) + reference tools	public	723	321	1	162	0	10	229	0
Qwen3 8B (stock, Q4_K_M, thinking off) + reference tools	public	723	282	0	289	0	22	130	0

How to read the rows:

• Pulse Local, rules is the app with "Better understanding" off: its own code and its tuned 0.6B model. Pulse Local 0.4.16, rules + 4B checker is the app as 0.4.16 shipped it with "Better understanding" on: the 4B checker (Qwen3-4B-Instruct-2507) read every message too, and the app showed a figure only when the two readings agreed; otherwise it said it was not sure. From 0.4.0 to 0.4.18 the 4B was the checker (from 0.4.1 on by default only above 16 GB of memory).
• Pulse Local 0.4.19, rules + checker is the app as it ships, "Better understanding" on (the default on every machine): the checker is a Qwen3-0.6B fine-tuned to read questions (PulseOS/pulse-local-checker-qwen3-0.6b, 397 MB, sha256 70f9ce49fc6397df0c460d4844928871eb851e03abeb63067db6a29857e59f59), called exactly as pulse-local/app/electron/llm/secondReader.ts calls it (the shipped variant). Its release claim is the v3 sealed run (PREREG-0419-v3sealed.md), not the tables above: on a sealed test of 387 conversations it had never seen, about 6 answers in 100 still showed a wrong figure, against about 23 in 100 without it (95% interval of the first figure 2.6% to 9.4%, households resampled). History: v2 sealed (2026-10-22, 340 cases, the 0.4.18 code, PREREG-small-checker-v2sealed.md): rules alone 73 wrong / 165 right, the 4B 17 / 103, the 0.6B checker (reader5) 16 / 140, 1.1 s against 11.7 s a turn.
• Stock models are Qwen3 0.6B, 4B and 8B (Q4_K_M, thinking off), each behind llama-server with the same system prompt and the same reference tools (systems/tools.mjs). They read parsed rows; Pulse Local reads the raw files (METHODOLOGY.md §7). They are not tuned for this task.
• No hosted model. BudgetBench runs and publishes no row for a paid hosted model of its own. Anyone may run one with their own key (SUBMIT.md) and send the row.
• Needs review rows are read by hand. Each run folder's ADJUDICATION.md gives the rulings and the adjudicated counts beside the grader's.
• Pulse Local's release claim is a sealed split. For 0.4.16 to 0.4.18 it was the sealed table above; for 0.4.19 it is v3 sealed (the row above). Its public rows are development scores.

Reproduce

From the folder that holds budgetbench/ (Node 24; Python 3 for check.mjs; Apple Silicon and llama.cpp's llama-server for the local models):

node budgetbench/check.mjs                         # frozen files, both oracles, the grader's 5,031 public checks: must pass
budgetbench/runner/serve-local.sh public           # the three stock Qwen3 rows, one llama-server at a time
node budgetbench/stats/report.mjs budgetbench/results/openai-compat-qwen3-8b/public/<run-id> --regrade
node budgetbench/stats/leaderboard.mjs <run folder> ...    # the tables above, regraded from the transcripts

The stock rows need the GGUF files in $MODELS (default app/models) under these names and SHA-256 sums:

row	file	source (Q4_K_M file)	SHA-256
Qwen3 0.6B	hf_unsloth_Qwen3-0.6B-Q4_K_M.gguf	unsloth/Qwen3-0.6B-GGUF	ac2d97712095a558e31573f62f466a3f9d93990898b0ec79d7c974c1780d524a
Qwen3 4B	hf_Qwen_Qwen3-4B-Q4_K_M.gguf	Qwen/Qwen3-4B-GGUF	7485fe6f11af29433bc51cab58009521f205840f5b4ae3a32fa7f92e8534fdf5
Qwen3 8B	Qwen3-8B-Q4_K_M.gguf	unsloth/Qwen3-8B-GGUF	120307ba529eb2439d6c430d94104dabd578497bc7bfe7e322b5d9933b449bd4

Each SHA-256 equals the repository's file on 2026-10-19. serve-local.sh writes the SHA-256 and the llama.cpp build of every stock run to results/logs/models.txt. Context 32,768, temperature 0.

Pulse Local rows run the app's own pipeline (systems/pulse-local.mjs), so they need the Pulse Local app source, which is not public:

node --experimental-strip-types --no-warnings budgetbench/runner/run.mjs --system pulse-local --variant rules  --split public --run-id <id>
node --experimental-strip-types --no-warnings budgetbench/runner/run.mjs --system pulse-local --variant reader --split public --run-id <id>

Anyone can still check every published row: each run folder holds every case's full transcript, and report.mjs --regrade recomputes every label from those transcripts with the frozen grader. Each inputs.json holds the SHA-256 of the cases, households, grader, oracle, runner, system file and (for Pulse Local) the app source.

Your own model: any OpenAI-compatible endpoint, with your own key in your own environment:

export MY_KEY=...        # your key; BudgetBench reads it from the process environment only and never writes it
BB_BASE=https://api.example.com/v1 BB_MODEL=<model id> BB_KEY=MY_KEY \
  node budgetbench/runner/run.mjs --system openai-compat --variant <name> --split public --run-id <id>
node budgetbench/stats/cost-estimate.mjs --price-in <$ per M> --price-out <$ per M>   # estimate first

See SUBMIT.md to send a row. The sealed split goes to no system under test through an API. Hosted Claude sessions wrote and audited it, and a Claude session ran and graded the v1 sealed runs (CONTAMINATION.md §6).

Files

file	contents
SPEC.md	the design: households, categories, oracle, grader, metrics, splits, systems
DATASHEET.md	how the data was made, what it covers, its known flaws
METHODOLOGY.md	the gold, the grader's rules, how each label is counted, the statistics, threats to validity
CONTAMINATION.md	the sealed split, what is exposed, what a submitter declares, what invalidates a result
SUBMIT.md	how to run your system and send a row
CHANGELOG.md	versions
cases/public/	the public cases (JSONL); viewer/public.jsonl is the same cases flattened for the Hub's viewer
households/public/	the public households: source/, view/, truth/
oracle/, grader/	the gold and the scorer
runner/, systems/, stats/	the runner, the systems, the reports
results/	every public run: each case's full transcript and label; results/SEALED-AGGREGATES.md
sealed.sha256	the SHA-256 of every sealed file, so anyone can check later that the sealed split did not change
SEALED-V2.md, sealed-v2.sha256	the v2 sealed split (2026-10-21, 340 cases on 6 new households): how it was made, its counts, and the SHA-256 of its 114 files
SEALED-V3.md, sealed-v3.sha256	the v3 sealed split (2026-10-22, 387 cases on 8 new households, for 0.4.19): how it was made, its counts, and the SHA-256 of its 149 files

Limits

• The households are synthetic. A score here is not a score on real ledgers.
• Most households are in US dollars. All questions are in English.
• The 566 writers' cases were written by Claude sessions acting as persons, not by real users. The other 494 public cases come from Pulse Local's earlier tests.
• 494 public cases are exposed to Pulse Local. Read its public rows with that in mind; its release claims come from the sealed split.
• Pulse Local reads the raw files; the model rows read parsed rows through tools.
• The grader reads text by rules. Every "needs review" row is ruled by a reviewer (for v1, a Claude session working for the maintainers). Every wrong row of the Pulse Local runs was read; the stock rows' wrong rows were read by the grader's reason, grouped, with a sample in full. Each run folder's ADJUDICATION.md says what was read.
• "Right after one tap" uses the answer key to choose the option. It is an upper bound.
• The team that makes Pulse Local made BudgetBench. Read Pulse Local's rows with that in mind too.

Licence

• Code (households generator, oracle, grader, runners, reports): Apache License 2.0. It is the licence of the Pulse Local model on Hugging Face, it is permissive, and it gives an explicit patent grant, which matters for code others will build into their own systems.
• Data (households, cases, golds, results): Creative Commons Attribution 4.0 (CC BY 4.0). Anyone may copy, change and use it, also commercially, if they credit BudgetBench.
• help/HELP.md is the text of Pulse Local's public help pages, included so that how-to cases can be graded. Its inclusion under CC BY 4.0 needs the owner's confirmation before publication (not yet confirmed).
• The sealed split is not released under any licence.

Citation

@misc{budgetbench2026,
  title        = {BudgetBench: Measuring Wrong Figures in Budgeting Chat Assistants},
  author       = {Grira, Nizar and {Pulse Local team}},
  year         = {2026},
  version      = {1.0.1},
  note         = {Public split of 723 cases on 12 synthetic households; sealed split of 337 cases on 6 households},
  howpublished = {\url{https://huggingface.co/datasets/PulseOS/budgetbench}}
}

Publishing (the owner's step)

node budgetbench/hf/export.mjs /tmp/budgetbench-hf     # the committed tree only; refuses any sealed content
hf upload PulseOS/budgetbench /tmp/budgetbench-hf . --repo-type dataset --private --commit-message "BudgetBench v1.0.1"

Contact: niz@pulseos.finance.

BudgetBench is a trademark of Pulse Browser, Inc.