Datasets:

trjfcgsgsaz
/

llm-bench

model stringclasses 16 values	task stringclasses 8 values	score float64 0 100	status stringclasses 3 values	deadline_scored bool 2 classes	seconds float64 0.31 900	timeout_seconds int64 300 1.2k	version stringclasses 6 values	reasoning_budget int64 -1 4.1k	response_cap int64 8.19k 32.8k	agent_profile stringclasses 3 values	backend stringclasses 2 values	temperature float64 0 1 ⌀	source stringlengths 26 48	anchor_score float64 9.59 100 ⌀	hard_score float64 18.3 100 ⌀	hard_correct_components float64 11 96 ⌀	hard_total_components float64 60 96 ⌀
flint	bugfix-history	100	completed	false	163.316	300	2.1.0	-1	32,768	editable-v1	llama.cpp	0	runs/batch-20261024-101521-856927	null	null	null	null
flint	bugfix-history	100	completed	false	133.228	300	2.1.0	-1	8,192	editable-v1	llama.cpp	null	runs/batch-20261024-095857-158353	null	null	null	null
flint	bugfix-reply-filter	100	completed	false	201.662	420	2.1.0	-1	32,768	editable-v1	llama.cpp	0	runs/batch-20261024-101521-856927	null	null	null	null
flint	bugfix-reply-filter	38.89	completed	false	87.722	420	2.1.0	-1	8,192	editable-v1	llama.cpp	null	runs/batch-20261024-095857-158353	null	null	null	null
flint	code-search	0	invalid_answer	false	45.27	540	1.1.0	-1	32,768	workflow-v1	llama.cpp	0	runs/batch-20261024-101521-856927	null	null	null	null
flint	code-search	0	invalid_answer	false	151.752	540	1.1.0	-1	8,192	workflow-v1	llama.cpp	null	runs/batch-20261024-095857-158353	null	null	null	null
flint	code-search	94.44	completed	false	233.049	540	1.1.0	4,096	32,768	workflow-v1	llama.cpp	0	runs/batch-20261024-110938-058652	null	null	null	null
flint	dialogue-customs	0	invalid_answer	false	367.917	540	2.0.0	-1	32,768	workflow-v1	llama.cpp	0	runs/batch-20261024-101521-856927	null	null	null	null
flint	dialogue-customs	0	invalid_answer	false	88.322	540	2.0.0	-1	8,192	workflow-v1	llama.cpp	null	runs/batch-20261024-095857-158353	null	null	null	null
flint	dialogue-customs	60.04	completed	false	55.181	540	2.0.0	4,096	32,768	workflow-v1	llama.cpp	0	runs/batch-20261024-110440-590321	83.53	54.17	52	96
flint	feature-trace	100	completed	false	33.505	300	1.0.0	-1	32,768	workflow-v1	llama.cpp	0	runs/batch-20261024-101521-856927	null	null	null	null
flint	feature-trace	100	completed	false	24.544	300	1.0.0	-1	8,192	workflow-v1	llama.cpp	null	runs/batch-20261024-095857-158353	null	null	null	null
flint	prose-characters	0	invalid_answer	false	373.472	540	2.0.0	-1	32,768	workflow-v1	llama.cpp	0	runs/batch-20261024-101521-856927	null	null	null	null
flint	prose-characters	0	invalid_answer	false	89.524	540	2.0.0	-1	8,192	workflow-v1	llama.cpp	null	runs/batch-20261024-095857-158353	null	null	null	null
flint	prose-characters	43.78	completed	false	56.233	540	2.0.0	4,096	32,768	workflow-v1	llama.cpp	0	runs/batch-20261024-110440-590321	78.88	35	21	60
flint	prose-transitions	0	invalid_answer	false	373.775	540	2.0.0	-1	32,768	workflow-v1	llama.cpp	0	runs/batch-20261024-101521-856927	null	null	null	null
flint	prose-transitions	0	invalid_answer	false	89.473	540	2.0.0	-1	8,192	workflow-v1	llama.cpp	null	runs/batch-20261024-095857-158353	null	null	null	null
flint	prose-transitions	33.76	completed	false	59.587	540	2.0.0	4,096	32,768	workflow-v1	llama.cpp	0	runs/batch-20261024-110938-058652	72.99	23.96	23	96
flint	regression-tests	80	completed	false	173.427	300	2.1.0	-1	32,768	workflow-v1	llama.cpp	0	runs/batch-20261024-101521-856927	null	null	null	null
flint	regression-tests	0	invalid_answer	false	84.419	300	2.1.0	-1	8,192	workflow-v1	llama.cpp	null	runs/batch-20261024-095857-158353	null	null	null	null
gemma12b	bugfix-history	100	completed	false	381.987	900	2.2.0	4,096	32,768	editable-v1	llama.cpp	null	runs/gemma12b-timeout-affected-20261030	null	null	null	null
gemma12b	bugfix-reply-filter	95.83	completed	false	403.832	420	2.1.0	4,096	32,768	editable-v1	llama.cpp	null	runs/batch-20261024-202455-522408	null	null	null	null
gemma12b	code-search	0	invalid_answer	false	370.168	900	1.2.0	4,096	32,768	workflow-v1	llama.cpp	null	runs/gemma12b-timeout-affected-20261030	null	null	null	null
gemma12b	dialogue-customs	79.64	completed	false	157.675	540	2.0.0	4,096	32,768	workflow-v1	llama.cpp	null	runs/batch-20261024-202455-522408	98.18	75	72	96
gemma12b	feature-trace	100	completed	false	132.636	300	1.0.0	4,096	32,768	workflow-v1	llama.cpp	null	runs/batch-20261024-202455-522408	null	null	null	null
gemma12b	prose-characters	0	invalid_answer	false	458.683	540	2.0.0	-1	16,384	workflow-v1	llama.cpp	1	runs/batch-20261024-220848-255959	null	null	null	null
gemma12b	prose-characters	0	invalid_answer	false	460.286	540	2.0.0	2,048	16,384	workflow-v1	llama.cpp	0	runs/batch-20261024-215805-976760	null	null	null	null
gemma12b	prose-characters	51.64	completed	false	157.057	900	2.1.0	4,096	32,768	workflow-v1	llama.cpp	null	runs/gemma12b-timeout-affected-20261030	91.53	41.67	25	60
gemma12b	prose-transitions	70.52	completed	false	170.784	540	2.0.0	4,096	32,768	workflow-v1	llama.cpp	null	runs/batch-20261024-202455-522408	98.45	63.54	61	96
gemma12b	regression-tests	50	completed	false	127.778	300	2.1.0	4,096	32,768	workflow-v1	llama.cpp	null	runs/batch-20261024-202455-522408	null	null	null	null
gemma26b	bugfix-history	100	completed	false	158.962	300	2.1.0	4,096	16,384	editable-v1	llama.cpp	0.1	runs/coding-v21-history-reused-20261018	null	null	null	null
gemma26b	bugfix-reply-filter	100	completed	false	126.121	420	2.1.0	4,096	16,384	editable-v1	llama.cpp	0.1	runs/coding-v21-filter-corrected-20261018	null	null	null	null
gemma26b	code-search	100	completed	false	253.176	540	1.1.0	4,096	32,768	baseline	llama.cpp	null	runs/batch-20261016-174527-094360	null	null	null	null
gemma26b	dialogue-customs	93.74	completed	false	77.66	540	2.0.0	4,096	16,384	workflow-v1	llama.cpp	0	runs/prose-v2-stable-context-20261017	93.68	93.75	90	96
gemma26b	feature-trace	92.5	completed	false	53.079	300	1.0.0	4,096	32,768	baseline	llama.cpp	null	runs/batch-20261016-174527-094360	null	null	null	null
gemma26b	prose-characters	46.05	completed	false	82.116	540	2.0.0	4,096	16,384	workflow-v1	llama.cpp	0	runs/prose-v2-stable-context-20261017	83.57	36.67	22	60
gemma26b	prose-transitions	57.99	completed	false	86.721	540	2.0.0	4,096	16,384	workflow-v1	llama.cpp	0	runs/prose-v2-stable-context-20261017	98.26	47.92	46	96
gemma26b	regression-tests	40	completed	false	49.779	300	2.1.0	4,096	32,768	baseline	llama.cpp	null	runs/batch-20261016-220740-860930	null	null	null	null
gemma31b	bugfix-history	100	completed	false	215.584	300	2.1.0	4,096	16,384	editable-v1	llama.cpp	0.1	runs/coding-v21-remaining-models-20261018	null	null	null	null
gemma31b	bugfix-reply-filter	97.22	completed	false	151.852	420	2.1.0	4,096	16,384	editable-v1	llama.cpp	0.1	runs/coding-v21-remaining-models-20261018	null	null	null	null
gemma31b	code-search	77.78	completed	false	440.808	540	1.1.0	4,096	32,768	baseline	llama.cpp	null	runs/batch-20261016-174527-094360	null	null	null	null
gemma31b	dialogue-customs	100	completed	false	137.886	540	2.0.0	4,096	16,384	workflow-v1	llama.cpp	0	runs/prose-v2-stable-context-20261017	100	100	96	96
gemma31b	feature-trace	92.5	completed	false	229.745	300	1.0.0	4,096	32,768	baseline	llama.cpp	null	runs/batch-20261016-174527-094360	null	null	null	null
gemma31b	prose-characters	77.93	completed	false	121.716	540	2.0.0	4,096	16,384	workflow-v1	llama.cpp	0	runs/prose-v2-stable-context-20261017	96.32	73.33	44	60
gemma31b	prose-transitions	82.77	completed	false	133.336	540	2.0.0	4,096	16,384	workflow-v1	llama.cpp	0	runs/prose-v2-stable-context-20261017	97.2	79.17	76	96
gemma31b	regression-tests	45.83	completed	false	130.982	300	2.1.0	4,096	32,768	baseline	llama.cpp	null	runs/batch-20261016-220740-860930	null	null	null	null
gemmae4b	bugfix-history	91.53	completed	false	23.694	300	2.1.0	4,096	16,384	editable-v1	llama.cpp	0.1	runs/coding-v21-history-reused-20261018	null	null	null	null
gemmae4b	bugfix-reply-filter	90.28	completed	false	55.784	420	2.1.0	4,096	16,384	editable-v1	llama.cpp	0.1	runs/coding-v21-filter-corrected-20261018	null	null	null	null
gemmae4b	code-search	0	completed	false	62.492	540	1.1.0	4,096	16,384	workflow-v1	llama.cpp	null	runs/batch-20261017-115811-508460	null	null	null	null
gemmae4b	code-search	0	completed	false	49.826	540	1.1.0	4,096	32,768	workflow-v1	llama.cpp	null	runs/batch-20261017-223439-281346	null	null	null	null
gemmae4b	dialogue-customs	84.52	completed	false	66.596	540	2.0.0	4,096	16,384	workflow-v1	llama.cpp	0	runs/prose-v2-e4b-20261017	85.12	84.38	81	96
gemmae4b	feature-trace	85	completed	false	29.3	300	1.0.0	4,096	16,384	workflow-v1	llama.cpp	null	runs/batch-20261017-115811-508460	null	null	null	null
gemmae4b	prose-characters	38.89	completed	false	70.752	540	2.0.0	4,096	16,384	workflow-v1	llama.cpp	0	runs/prose-v2-e4b-20261017	87.78	26.67	16	60
gemmae4b	prose-transitions	48.04	completed	false	58.839	540	2.0.0	4,096	16,384	workflow-v1	llama.cpp	0	runs/prose-v2-e4b-20261017	98.54	35.42	34	96
gemmae4b	regression-tests	36.67	completed	false	28.85	300	2.1.0	4,096	16,384	workflow-v1	llama.cpp	null	runs/batch-20261017-115811-508460	null	null	null	null
lfm24b	bugfix-history	93.22	completed	false	73.105	300	2.1.0	4,096	16,384	editable-v1	llama.cpp	0.1	runs/coding-v21-remaining-models-20261018	null	null	null	null
lfm24b	bugfix-reply-filter	87.5	completed	false	173.178	900	2.2.0	4,096	16,384	editable-v1	llama.cpp	0.1	runs/lfm24b-timeout-affected-20261030	null	null	null	null
lfm24b	code-search	100	completed	false	132.562	540	1.1.0	4,096	32,768	baseline	llama.cpp	null	runs/batch-20261016-174527-094360	null	null	null	null
lfm24b	code-search	88.89	completed	false	118.411	540	1.1.0	4,096	32,768	workflow-v1	llama.cpp	null	runs/batch-20261017-223748-240819	null	null	null	null
lfm24b	dialogue-customs	44.27	completed	false	44.018	540	2.0.0	4,096	16,384	workflow-v1	llama.cpp	0	runs/prose-v2-stable-context-20261017	67.2	38.54	37	96
lfm24b	feature-trace	77.5	completed	false	56.09	300	1.0.0	4,096	32,768	baseline	llama.cpp	null	runs/batch-20261016-174527-094360	null	null	null	null
lfm24b	prose-characters	21.25	completed	false	46.522	540	2.0.0	4,096	16,384	workflow-v1	llama.cpp	0	runs/prose-v2-stable-context-20261017	32.92	18.33	11	60
lfm24b	prose-transitions	34.67	completed	false	48.073	540	2.0.0	4,096	16,384	workflow-v1	llama.cpp	0	runs/prose-v2-stable-context-20261017	77.53	23.96	23	96
lfm24b	regression-tests	50	completed	false	34.41	300	2.1.0	4,096	32,768	baseline	llama.cpp	null	runs/batch-20261016-220740-860930	null	null	null	null
lfm3	bugfix-history	84.75	completed	false	238.507	300	2.1.0	4,096	32,768	editable-v1	llama.cpp	null	runs/lfm3-1.7b-q8-20261028	null	null	null	null
lfm3	bugfix-reply-filter	65.28	completed	false	154.103	420	2.1.0	4,096	32,768	editable-v1	llama.cpp	null	runs/lfm3-1.7b-q8-20261028	null	null	null	null
lfm3	code-search	0	completed	false	414.222	540	1.1.0	4,096	32,768	workflow-v1	llama.cpp	null	runs/lfm3-1.7b-q8-20261028	null	null	null	null
lfm3	dialogue-customs	35.18	completed	false	35.958	540	2.0.0	4,096	32,768	workflow-v1	llama.cpp	null	runs/lfm3-1.7b-q8-20261028	42.58	33.33	32	96
lfm3	feature-trace	100	completed	false	108.847	300	1.0.0	4,096	32,768	workflow-v1	llama.cpp	null	runs/lfm3-1.7b-q8-20261028	null	null	null	null
lfm3	prose-characters	23.39	completed	false	37.21	540	2.0.0	4,096	32,768	workflow-v1	llama.cpp	null	runs/lfm3-1.7b-q8-20261028	23.61	23.33	14	60
lfm3	prose-transitions	29.83	completed	false	33.255	540	2.0.0	4,096	32,768	workflow-v1	llama.cpp	null	runs/lfm3-1.7b-q8-20261028	36.63	28.12	27	96
lfm3	regression-tests	40	completed	false	26.596	300	2.1.0	4,096	32,768	workflow-v1	llama.cpp	null	runs/lfm3-1.7b-q8-20261028	null	null	null	null
museglimmer30b	bugfix-history	100	completed	false	499.628	900	2.2.0	-1	32,768	editable-v1	llama.cpp	null	runs/museglimmer30b-timeout-policy-20261030	null	null	null	null
museglimmer30b	bugfix-history	100	completed	false	406.765	1,200	2.1.0	4,096	32,768	editable-v1	llama.cpp	null	runs/museglimmer30b-history-20min-20261030	null	null	null	null
museglimmer30b	bugfix-reply-filter	98.61	completed	false	607.907	900	2.2.0	-1	32,768	editable-v1	llama.cpp	null	runs/museglimmer30b-timeout-policy-20261030	null	null	null	null
museglimmer30b	bugfix-reply-filter	98.61	timeout	true	420.013	420	2.1.0	4,096	32,768	editable-v1	llama.cpp	null	runs/museglimmer30b-remaining-20261030	null	null	null	null
museglimmer30b	code-search	94.44	completed	false	351.098	900	1.2.0	-1	32,768	workflow-v1	llama.cpp	null	runs/museglimmer30b-timeout-policy-20261030	null	null	null	null
museglimmer30b	code-search	100	completed	false	256.582	540	1.1.0	4,096	32,768	workflow-v1	llama.cpp	null	runs/museglimmer30b-rest-20261030	null	null	null	null
museglimmer30b	dialogue-customs	99.37	completed	false	202.562	1,200	2.0.0	-1	32,768	workflow-v1	llama.cpp	null	runs/museglimmer30b-prose-low-20min-20261030	96.84	100	96	96
museglimmer30b	dialogue-customs	0	timeout	false	540.008	540	2.0.0	4,096	32,768	workflow-v1	llama.cpp	null	runs/museglimmer30b-rest-20261030	null	null	null	null
museglimmer30b	feature-trace	100	completed	false	140.89	300	1.0.0	-1	32,768	workflow-v1	llama.cpp	null	runs/museglimmer30b-no-budget-remaining-20261030	null	null	null	null
museglimmer30b	prose-characters	70.36	completed	false	144.389	1,200	2.0.0	-1	32,768	workflow-v1	llama.cpp	null	runs/museglimmer30b-prose-low-20min-20261030	91.81	65	39	60
museglimmer30b	prose-transitions	99.03	completed	false	556.604	1,200	2.0.0	-1	32,768	workflow-v1	llama.cpp	null	runs/museglimmer30b-transitions-medium-20261030	99.31	98.96	95	96
museglimmer30b	regression-tests	55	completed	false	208.871	300	2.1.0	4,096	32,768	workflow-v1	llama.cpp	null	runs/museglimmer30b-regression-20261030	null	null	null	null
nanbeige43b	regression-tests	0	invalid_answer	false	0.314	1,200	2.1.0	-1	32,768	workflow-v1	llama.cpp	null	runs/nanbeige43b-regression-unlimited-20261030	null	null	null	null
nemo74b	bugfix-history	100	completed	false	137.838	300	2.1.0	4,096	16,384	editable-v1	llama.cpp	0.1	runs/coding-v21-remaining-models-20261018	null	null	null	null
nemo74b	bugfix-reply-filter	100	completed	false	258.684	420	2.1.0	4,096	16,384	editable-v1	llama.cpp	0.1	runs/coding-v21-remaining-models-20261018	null	null	null	null
nemo74b	code-search	88.89	completed	false	170.625	540	1.1.0	4,096	32,768	baseline	llama.cpp	null	runs/batch-20261016-174527-094360	null	null	null	null
nemo74b	dialogue-customs	89.29	completed	false	107.498	540	2.0.0	4,096	16,384	workflow-v1	llama.cpp	0	runs/prose-v2-stable-context-20261017	92.26	88.54	85	96
nemo74b	feature-trace	92.5	completed	false	34.255	300	1.0.0	4,096	32,768	baseline	llama.cpp	null	runs/batch-20261016-174527-094360	null	null	null	null
nemo74b	prose-characters	46.49	completed	false	113.104	540	2.0.0	4,096	16,384	workflow-v1	llama.cpp	0	runs/prose-v2-stable-context-20261017	92.47	35	21	60
nemo74b	prose-transitions	68.08	completed	false	124.268	540	2.0.0	4,096	16,384	workflow-v1	llama.cpp	0	runs/prose-v2-stable-context-20261017	98.74	60.42	58	96
nemo74b	regression-tests	50	completed	false	87.729	300	2.1.0	4,096	32,768	baseline	llama.cpp	null	runs/batch-20261016-220740-860930	null	null	null	null
nemoe4b	bugfix-history	69.49	completed	false	25.495	300	2.1.0	4,096	32,768	editable-v1	llama.cpp	0	runs/batch-20261024-114419-641794	null	null	null	null
nemoe4b	bugfix-reply-filter	38.89	completed	false	89.723	420	2.1.0	4,096	32,768	editable-v1	llama.cpp	0	runs/batch-20261024-114419-641794	null	null	null	null
nemoe4b	code-search	0	completed	false	95.682	540	1.1.0	4,096	32,768	workflow-v1	llama.cpp	0	runs/batch-20261024-114419-641794	null	null	null	null
nemoe4b	dialogue-customs	61.08	completed	false	90.075	540	2.0.0	4,096	32,768	workflow-v1	llama.cpp	0	runs/batch-20261024-114419-641794	9.59	73.96	71	96
nemoe4b	feature-trace	70	completed	false	27.347	300	1.0.0	4,096	32,768	workflow-v1	llama.cpp	0	runs/batch-20261024-114419-641794	null	null	null	null
nemoe4b	prose-characters	40.54	completed	false	93.429	540	2.0.0	4,096	32,768	workflow-v1	llama.cpp	0	runs/batch-20261024-114419-641794	89.38	28.33	17	60
nemoe4b	prose-transitions	46.89	completed	false	99.336	540	2.0.0	4,096	32,768	workflow-v1	llama.cpp	0	runs/batch-20261024-114419-641794	96.94	34.38	33	96

End of preview. Expand in Data Studio

Local model benchmarks

A modular benchmark suite for local models running through Pi with llama.cpp or vLLM, with deterministic scoring and short, timed tasks.

How to read these scores: Each task score comes from a single attempt, not an average over repeated trials; some tasks have multiple attempts and the dashboard selects the highest compatible score. Models were run with different reasoning budgets, output limits, temperatures, and time limits. I chose these tasks because they matter to me personally. The results describe performance on this particular set of tasks, not a model's overall intelligence or value.

Dashboard overview

Highest compatible score per model and task, across recorded settings; newest attempt wins ties. Prose averages three tests; coding weights its four families equally. These best recorded results mix launch settings, task deadlines, and numbers of attempts; use the interactive dashboard to filter settings and inspect individual results. Runtime belongs to the selected attempt and excludes model loading.

The October 30, 2026 timeout policy gives six longer tasks 900 seconds by default and permits explicit overrides up to 1,200 seconds. Their prompts, fixtures, and scoring stayed the same, so the dashboard accepts the older deadline-only fingerprints listed in the comparison configuration. The initial exclusion inventory records what the stricter filter had hidden; targeted rerun results cover prior timeout cases except Nanbeige. Earlier reports and the FLINT snapshot below remain historical results.

Browse the underlying scores · Download the interactive dashboard

Published results

Browse the scores online or download the comparison CSV. The dataset viewer's train split is a table of evaluation results, not model training examples. Each row is the highest compatible result for a model, task, and settings profile; newest attempt wins ties and profiles remain separate.

Historical snapshot: FLINT UD-Q4_K_XL, tested October 24, 2026 with llama.cpp at temperature 0.

Task	Score / 100	Reasoning budget	Output cap per turn
History repair	100.00	Unlimited	32,768
Reply-filter repair	100.00	Unlimited	32,768
Code search	94.44	4,096	32,768
Dialogue understanding	60.04	4,096	32,768
Feature tracing	100.00	Unlimited	32,768
Character understanding	43.78	4,096	32,768
Scene transitions	33.76	4,096	32,768
Regression-test design	80.00	Unlimited	32,768

See the FLINT report and run provenance. Earlier attempts remain in the imported snapshots; this table uses the latest temperature-0 attempt per task. The targeted code-search and prose reruns use reasoning budget 4096; the other tasks use unlimited reasoning from the initial full run. This summary mixes settings; the viewer and dashboard retain them separately. These are single attempts rather than repeated-run averages.

Reasoning budgets are shown as the specified token limit or Unlimited (no separate reasoning cap). Runs with no budget override used the uncapped server default. JSON and CSV encode Unlimited as -1; original imported run records retain the requested value, including null when omitted. Unlimited reasoning still respects the output cap and task deadline. All numeric budgets and caps are in tokens. The output cap includes reasoning plus the final answer and applies per turn, not per task. The reasoning budget is a separate cap per response; thinking: off is a separate Pi setting.

For comparison, the current Swift 1.5 (swift27b, Q4_K_L) records use these limits:

Tasks	Reasoning budget	Output cap per turn
History and reply-filter repairs	4,096	16,384
Dialogue, character, and scene understanding	4,096	16,384
Code search, feature tracing, and regression-test design	4,096	32,768

These limits differ from Flint’s selected results, so the scores do not represent a comparison with matched token limits. The dashboard displays reasoning budgets per task and shows the output limit only when that attempt is confirmed to have hit it. Raw data and CSV exports retain every configured cap; settings profiles remain separate.

The interactive HTML dashboard is included for local viewing: download/clone the repository and open dashboard.html. Hugging Face displays the scores through its dataset viewer and this card; HTML files are available to download.

Requirements

Python 3.9+; no Python packages required.
Node.js 22+ for JavaScript grading.
An installed Pi agent and working llama.cpp or vLLM launch scripts.

Quick start

Run all benchmarks for one model:

python run_models.py new-model.sh --reasoning-budget 4096 --max-tokens 32768

Select tasks or compare models sequentially:

python run_models.py gemma26b.sh swift27b.sh --tasks regression-tests --reasoning-budget 4096 --max-tokens 32768
python run_models.py new-model.sh --tasks bugfix-reply-filter dialogue-customs
python run_models.py --list
python run_models.py new-model.sh --dry-run
python run_models.py ../vllm/serve-dory.sh --reasoning-budget 4096 --max-tokens 16384

Script names resolve from the current directory or ../llama.cpp; explicit paths also work. Scripts run through Bash from their own directory and must keep the server in the foreground. --dry-run shows launch adaptations without starting servers.

Normal runs stop existing llama-server processes, launch each model in turn, and stop it afterward. Pi uses a fresh, isolated configuration for each model.

Run options

Option	Purpose
`--tasks ID ...`	Select benchmarks; default: all discovered tasks
`--script-dir PATH`	Change the launch-script search directory
`--backend NAME`	`auto` (default), `llama.cpp`, or `vllm`
`--temperature N`	Override temperature on every API request; other sampling settings stay unchanged
`--reasoning-budget N`	Server reasoning cap per response; `-1` = unlimited, `0` = end reasoning immediately
`--max-tokens N`	Response cap, including reasoning; default: 8192
`--agent-profile NAME`	`auto` (default): `editable-v1` for source editing, `workflow-v1` otherwise; explicit profiles override selection
`--timeout N`	Override whole-task timeout, up to 1,200 seconds
`--output PATH`	Write to a new results directory
`--server-url URL`	Override server URL detection, e.g. `http://127.0.0.1:9931`

Sampling settings come from each launch script. Explicit script flags override the runner's reasoning-budget and port environment settings. Without --reasoning-budget, the server's own setting applies.

Default deadlines are 900 seconds for the two repairs, code search, and three prose tasks; feature tracing and regression-test design remain at 300 seconds. The wall-clock limit is a fallback for stalled runs, and results always record actual runtime and timeout status.

The runner requests port 8080 and detects the actual port from startup logs. Context size comes from /props. See python run_models.py --help for further options.

vLLM scripts containing vllm serve are detected automatically. They keep their own port settings; context size comes from /v1/models, and --reasoning-budget is sent as thinking_token_budget on every request. A budget of -1 leaves vLLM thinking uncapped. vLLM uses its native reasoning-end marker rather than llama.cpp's budget message. Choose a response cap below the model's context limit to leave room for input. Results and comparison settings record the backend separately.

The default auto selection uses editable-v1 for bugfix-history and bugfix-reply-filter. Other tasks use workflow-v1, including schema-constrained tasks. --dry-run reports the resolved task_profiles; individual results record the actual profile used. Set --agent-profile workflow-v1 to reproduce the previous JSON submission workflow for coding tasks.

The workflow-v1 profile explains the available tools and how to submit an answer. A progress extension warns after three calls with the same arguments and unchanged results, then disables tools after five calls without new evidence. New evidence resets the repetition counter. It also warns near the task deadline and reserves the final 15% (up to 45 seconds) for submission after a tool result. The normal wall-clock timeout still applies, including during reasoning.

Results record the agent profile, exact system prompt, extension hash, and intervention counts; progress-audit.jsonl records individual reminders. The comparison report keeps baseline and workflow results in separate settings profiles. To reproduce the original interaction protocol:

python run_models.py new-model.sh --agent-profile baseline --reasoning-budget 4096 --max-tokens 32768

For source-editing exercises, editable-v1 uses Pi's stock system prompt and native read/edit/write/search/Bash tools on a fresh copy of the input. Saved modules are graded directly; the final message needs no JSON. On Linux, Bubblewrap isolates shell commands from the hidden grader and network. File tools and Bash share a private /tmp for scratch tests; only declared modules are graded. The original fixtures stay untouched. This profile has no loop monitor and stays separate in the comparison report. It is selected automatically for applicable tasks; it can also be requested explicitly:

python run_models.py lfm24b.sh --agent-profile editable-v1 --tasks bugfix-history bugfix-reply-filter --temperature 0.1 --reasoning-budget 4096 --max-tokens 32768 --reasoning-budget-message 'Thinking budget exceeded. Complete the file edits and give a brief final summary.'

Change --temperature to repeat the same tasks at another temperature. Ordinary timeouts score zero. The current coding repair tasks opt into grading source files saved by the deadline, keeping timeout status and elapsed time separate. Other editable tasks retain a diagnostic grade for saved files without changing their timeout score.

Benchmarks

ID	Task
`code-search`	Locate six related settings/functions in a pinned SillyTavern codebase
`bugfix-reply-filter`	Repair matching, validated settings, distinct counts, and atomic persistence across three modules (900 s)
`bugfix-history`	Select whole conversation turns with mandatory system messages, reserved costs, overhead, and count limits (900 s)
`dialogue-customs`	Attribute live, quoted, recorded, and synthetic speech; track roles, clothing, channels, and presence
`prose-transitions`	Reconstruct distributed event order and physical locations across all orders supported by the text
`prose-characters`	Track private observations and beliefs at one to four levels; interpret indirect emotions and motives
`feature-trace`	Locate a setting's flow and predict eight outcomes
`regression-tests`	Design up to 12 API test cases that detect hidden defects

The coding v2.2 repair tasks retain the v2.1 prompts and rubrics with 900-second deadlines for each task. Their combined wall-clock cap is 1,800 seconds. They use saved source files by default: the final message is not graded for JSON or prose format. The filter has 72 equally weighted checks (24 matching, 24 settings, 24 session/persistence); history has 59 checks. Harder requirements target interactions and incomplete fixes, with partial credit per behavior. The former tasks and grading dependencies are frozen in coding v1; their scores are not comparable to v2. Prior v2.1 results remain in the current dashboard because only the deadlines changed. Rebuild held-out cases with python development/build_coding_v2.py. The filter prompt requires case-insensitive matching and use of distinctWords in the rejection caller. Models receive identical requirements; calibration attempts with the earlier ambiguous prompt are excluded.

Scoring

Scores range from 0–100. Results also record completion status, format validity, elapsed time, and token usage. Active task defaults are 300 or 900 seconds; model loading has a separate timeout.

Repairs and features: execute submitted CommonJS code against held-out behavior and regression checks.
Prose v2: each of the three tasks retains its v1 material as a frozen anchor (20% of the score), followed by a substantial hard section (80%). Each hard answer component receives independent credit. Hard-section citations are diagnostic and cannot lower an otherwise correct score. The dialogue and transition sections contain 96 hard components each; character understanding contains 60. The hard section is presented first so the reasoning budget is spent on the higher-weight material. Use --temperature 0 for v2 comparisons; calibration found substantial native-temperature variation on the harder cases. Temperature remains part of each recorded settings profile. Tools-free runs use a fixed working directory because Pi includes it in the system context; changing the output folder therefore leaves model messages unchanged.
Frozen anchors: dialogue keeps the character/speaker/presence rubric; the two reading anchors keep their 24 multiple-choice questions, balanced answer positions/length ranks, and 75% choice/25% evidence rubric.
Code search: score symbols, files, and bounded line citations. Function citations must cover the declaration anchor.
Test design v2.1: independently check each case against the correct API and ten faulty versions. Score = 100 × valid-case fraction × defect coverage. Only correct, nonduplicate cases can detect defects; a bad case loses partial credit.

Test design and all three prose tasks use schema-constrained JSON, with the implementation supplied in the prompt and tools disabled. The schema controls structure; the graders check correctness, evidence, and test coverage. Other tasks allow source inspection within their input directories. This tool boundary and the bounded JavaScript evaluator are not OS sandboxes.

Schema-constrained tasks require strict JSON. Unconstrained tasks, including code-search, can recover one unambiguous JSON object from surrounding prose, recording the format violation separately without penalizing content scores. Multiple candidate objects remain invalid. Unscorable answers and ordinary timeouts receive zero. The two coding repair tasks grade verified source files at the fixed deadline, even when the agent has not finalized; these results retain timeout status and a deadline_scored marker. Token totals can omit unfinished responses.

The active prose family has three v2 tasks: dialogue, scene transitions, and character understanding. A v2 submission contains challenge and anchor objects; each hard question returns an ordered values list using its public vocabulary. The temporal/spatial answer key is checked against a partial-order calculator, and the nested-belief key against a separate belief-tree update model. The original prose pilot and two redundant dialogue scenes are preserved in retired-tasks, outside default discovery. Historical reports keep their original scores; the current dashboard excludes retired tasks and requires both new reading tasks for a complete prose-family score. The v1 versions are preserved byte-for-byte in archives/prose-v1/, and the previous report still replays against those fingerprints. These tests measure comprehension, not creative-writing quality.

To run the prose family:

python run_models.py new-model.sh --tasks prose-transitions dialogue-customs prose-characters --temperature 0 --reasoning-budget 4096 --max-tokens 16384

The hard passages and gold data are reproducibly authored in development/build_hard_prose.py. Run python development/build_hard_prose.py to rebuild challenge_cases.json, challenge_gold.json, the combined prompt, and its schema. Review narrative/answer-key agreement and bump the task version for any content or rubric change. cases.json, gold.json, and anchor_prompt.txt preserve the earlier anchor. Verification rejects stale generated material; tests also check source reproducibility and independently worked belief/ordering cases.

Compare models

Open dashboard.html in the repository root to view the latest generated report. To refresh it first:

python compare.py --open

Dashboard files have separate roles:

dashboard.html — the entry point to open in your browser.
results/comparison/ — generated report, data, exports, and imported results.
templates/dashboard.html.template — source template used by the report generator.
results/comparison/figures/ — dataset-card snapshots exported from the dashboard.

The dark dashboard shows separate prose and coding leaderboards. Prose opens by default; switch between Prose, Coding, and All tests for the detailed score/runtime plot and heatmap. Each leaderboard uses all tests in its domain and the selected models/settings; test filters affect the detailed views. Coding weights its four families equally, while prose averages its three tests. Click cells for failures and provenance, or export the selected results as CSV.

New model batches and python compare.py automatically refresh all three dataset-card images alongside the report. Install the image exporter once on each machine:

python development/export_dashboard.py --setup

This installs Playwright and Chromium inside the ignored .venv/ directory. Report refreshes use that environment automatically. No Git commits or pushes happen automatically; commit and push the updated results and images when ready to publish. If image export fails, benchmark results and the report remain saved; retry with python compare.py. Use python compare.py --no-figures to rebuild only HTML/JSON/CSV. For a manual image export, run .venv/bin/python development/export_dashboard.py.

Future runs refresh it automatically. Run a new model normally, then reload the browser tab. New model names are discovered without configuration changes. Use Only tests shared by selected models when comparing partial runs.

compare.py imports completed runs from runs/, keeps compact copies in results/comparison/imports/, and selects the highest compatible score per model, task, and settings profile, breaking score ties by newest attempt. The dashboard then selects the highest score across the selected profiles. Current benchmark fingerprints are required. Generated HTML, JSON, and CSV snapshots and imported results are versioned so published results are available immediately. Regenerate with python compare.py before committing new results.

To import an existing run stored elsewhere:

python compare.py --add /path/to/batch-results --open

Optional display names and family groups live in comparison configuration. The All tests detailed view weights all five families equally; missing tests leave that combined mean unranked. The main leaderboards keep prose and coding separate. Older reports below use different aggregates and test-writing versions.

Results

New runs go to runs/batch-TIMESTAMP/. Each contains summary.json, per-task answers and scores, raw Pi events, and launch/server metadata. Raw runs are Git-ignored; curated reports in results/ are tracked.

Compare task scores, failure modes, and runtime. These reports use single attempts with each model's launch settings; repeat runs before drawing general rankings. The retired published-story pilot may appear in training data. Test-writing v1 and test-design v2 use different protocols and scoring. Earlier coding-task versions are preserved in archives/; retired prose tasks are in retired-tasks/.

Add a benchmark

python bench.py new my-test --title "My new test"
# Edit tasks/my-test/: benchmark.json, prompt.txt, score.py, and reference data.
python bench.py verify
python run_models.py new-model.sh --tasks my-test

Tasks are discovered automatically from tasks/*/benchmark.json. See the manifest and scorer contract and manifest JSON Schema. Use --tasks-dir PATH for an alternative task collection.

Tools-free tasks can set response_schema to a JSON Schema file to request constrained output from a compatible server. Results record the schema and request-hook hashes. Bump the task version when changing its prompt, rubric, or interaction protocol. Agent-wide workflow changes are versioned separately through --agent-profile.

Use an already configured Pi model

python bench.py run prose-characters --provider YOUR_PROVIDER --model YOUR_MODEL --output runs/manual-characters-1
python bench.py score prose-characters answer.json

Verify

python bench.py verify
python -m unittest discover -s tests

Checks use fixtures and fake APIs without loading models or stopping a real server.

The repository includes the pinned SillyTavern source snapshot, its license, and file hashes. Git preserves exact bytes so fixture hashes survive checkout. To recreate a missing snapshot from the pinned commit:

python bench.py prepare --source ../SillyTavern

Model weights and launch scripts are supplied separately. Raw runs, caches, and local configuration are Git-ignored.

Downloads last month: 179