axonis22/axonis-judge-papers

Axonis Pty Limited Judge Papers (retired)

These are nine small, frozen test papers (136 items in total) that Axonis Pty Limited wrote and checked. They were used to test language models running locally on owned hardware in Australia, through llama.cpp and vLLM. The papers are now retired, because the models tested scored at or near the ceiling on them. Papers, keys, scoring scripts and checks: Axonis Pty Limited. Curated by Paul Buckingham.

Paper	Items	Content
CLIFF-16	16	HTML apps, Python, maths, physics, text locks, multi-turn poison
CLIFF-24	24	harder follow-up: Python that is executed, maths, physics, long-context lookup, JSON and lipogram locks, HTML, multi-turn poison
IQ-12	12	short IQ-style puzzles (not clinical, not a measure of anything about people)
HOLE-16	16	landmine HTML, Python, maths, physics, text, multi-turn
WEAK-20	20	HTML, executed Python, text constraints, 4 vision counting items, multi-turn
MATH-12 / MATH-HARD-12	12 each	short-answer maths
PHYS-12 / PHYS-HARD-12	12 each	short-answer physics

Results (first try, manual check)

Paper	Items	Siona	Sirin	Qwen 3.6
CLIFF-16	16	16	16	16
CLIFF-24	24	24	23	19
IQ-12	12	12	12†	12
HOLE-16	16	16	pending	15
WEAK-20	20	20	20	19
MATH-12	12	12	12†	12
PHYS-12	12	12†	12†	12
MATH-HARD-12	12	12†	12†	12
PHYS-HARD-12	12	12†	12†	12

• Siona: Muse 2 (muse-glimmer-30b) on llama.cpp, dual Tesla V100 32GB server.
• Sirin: Muse (muse-glimmer-30b) on llama.cpp, dual Tesla P100 16GB server.
• Qwen 3.6: Qwen3.6-35B-A3B-AWQ-4bit on vLLM, ASUS, dual RTX 3090. About 71 to 126 tok/s across the papers.
• † No manual check recorded; the mechanical score is shown, and the scorer found nothing to overturn.
• "pending" means the paper has not been sat yet.
• Mechanical scores are kept in results/ next to the manual check. They are lower in places, mostly because of scorer false negatives. Qwen 3.6's mechanical scores are much lower on CLIFF-16 (5), HOLE-16 (5) and WEAK-20 (14): the vLLM server had no reasoning parser, so Qwen's thinking leaked into the answer text, and the scorer read it all. The manual check counts only the text after </think>. Qwen's CLIFF-16, IQ-12, HOLE-16 and WEAK-20 results are runner-verified, not fingerprint-verified.
• Older Qwen 3.8 27B results from September 2026 are in report.md and results/.

Apple open models (mechanical grading)

On 25 Oct 2026 three of Apple's open-weights models sat the nine papers on one local server (Dell R730xd, dual GTX 1070 from 2016, llama.cpp). "Pippa" is the local deployment name. Quants: SimpleSD-4B-instruct Q8_0, SimpleSD-30B-instruct i1-Q3_K_S (ctx 12288 to fit 2 x 8 GB; about 33 tok/s), OpenELM-3B-Instruct Q8_0 (ctx 2048). Apple's open models trail Muse on our nine papers.

Scores are mechanical, first try only; no manual check was done for the Apple models. WEAK-20 is scored on W01 to W14 only (the runner stops at W15 because the vision images are withheld), so every total is out of 130. For a like-for-like comparison the other deployments are shown as mechanical / manual check on the same 130 items. Their WEAK-20 W01 to W14 figures are worked out from the item-level fails recorded in results/.

Paper	Items	SimpleSD-4B	SimpleSD-30B	OpenELM-3B	Siona (Muse)	Sirin (Muse)	Qwen 3.6
CLIFF-16	16	10	9	1	16 / 16	15 / 16	5 / 16
CLIFF-24	24	15	15	0	24 / 24	23 / 23	19 / 19
HOLE-16	16	12	6	0	15 / 16	–	5 / 15
IQ-12	12	8	11	0	12 / 12	12 / 12†	12 / 12
WEAK-20 (W01–W14)	14	12	9	0	9 / 14	8 / 14	9 / 14
MATH-12	12	12	12	0	11 / 12	12 / 12†	12 / 12
PHYS-12	12	12	12	0	12 / 12†	12 / 12†	12 / 12
MATH-HARD-12	12	6	6	0	12 / 12†	12 / 12†	12 / 12
PHYS-HARD-12	12	11	12	0	12 / 12†	12 / 12†	12 / 12
Total	130	98	92	1	123 / 130	106 / 113 of 114 ‡	98 / 124

• † No manual check recorded; the mechanical score is used for the manual column.
• ‡ Sirin did not sit HOLE-16, so its basis is 114 items.
• Siona and Sirin both run Muse Glimmer 30B (Siona: Q4_K_M, dual Tesla V100 32GB; Sirin: dual Tesla P100 16GB). Qwen 3.6 is Qwen3.6-35B-A3B-AWQ-4bit on vLLM, dual RTX 3090; its low mechanical CLIFF-16 and HOLE-16 scores come from leaked thinking text (see above).
• Possible format-only regrades, not applied: SimpleSD-4B MATH-HARD-12 MH09 (answered as a LaTeX fraction equal to the key), giving 99/130; SimpleSD-30B CLIFF-16 P02, CLIFF-24 P02, HOLE-16 M03, P02 and P01 (trailing full stop, float form or unit on a correct value), giving 97/130.
• OpenELM-3B with a 2048-token context mostly hit the token limit or echoed the prompt template.

Nine-paper re-runs, auto-marked only (31 Oct – 2 Nov 2026)

New sits with one shared harness: same runners, prompts and scoring scripts, temp 0.2, thinking off (enable_thinking=false), QCTX=32768, first try only, and marked only by each paper's own script (no hand check). All 136 items, including WEAK-20 W15–W18.

Paper	Items	Muse Glimmer 30B (Siona, re-run)	Siropia: Qwen3.8-Flash-Next IQ3_S (Siona, Strata)	Nemotron 3.5 Super 120B	Qwen 3.8 27B IQ4 (Dell)	Qwen 3.6 35B (ASUS)	Ornith 1.5 122B
CLIFF-16	16	15	15	14	12	13	11
IQ-12	12	12	12	12	12	12	11
HOLE-16	16	14	14	14	13	14	14
WEAK-20	20	15	18	12	12	14	10
MATH-12	12	11	7	12	12	9	12
PHYS-12	12	12	12	11	12	12	11
MATH-HARD-12	12	11	9	12	10	9	12
PHYS-HARD-12	12	12	12	12	12	12	12
CLIFF-24	24	23	22	22	21	19	19
Total	136	125	121	121	116	114	112
Mean gen speed in the sit		53.2 tok/s	86.3 tok/s	12.0 tok/s	–	–	7.6 tok/s

• Hand-checked vs auto-marked. Muse's 25 Oct sit on Siona scored 136/136 after a hand check. The 125/136 above is the auto-marked re-run (1 Nov 2026) and is the like-for-like number against the other columns, all of which are auto-marked.
• Models: Muse Glimmer 30B Q4_K_XL (llama.cpp, 2x Tesla V100 32GB, DFlash draft n-max 16, layer split, q8_0 KV, ctx 131072); Siropia = Qwen3.8-Flash-Next GSQ-RCO IQ3_S on the Strata engine (experimental CUDA 12 sm_70 build, 2x Tesla V100 32GB on Siona, auto layer split, all experts resident, spec 4 + MTP, int8 KV, ctx 131072; Muse was paused for the sit and restored afterwards); Nemotron-3.5-Super-VL-120B-A12B UD-Q4_K_XL (llama.cpp, 2x V100); Ornith-1.5-122B Q8_0 (llama.cpp, 2x V100); Qwen 3.8 27B UD-IQ4_XS (llama.cpp, Dell R730xd 2x GTX 1070); Qwen3.6-35B-A3B-AWQ-4bit (vLLM, ASUS 2x RTX 3090).
• Speeds are completion tokens / call time from the harness (Muse 44019 tokens / 827.8 s; Siropia 66105 / 766.1 s, per paper 70.8–98.5 tok/s; Nemotron 28060 / 2340.0 s; Ornith 30561 / 4032.7 s).
• Muse re-run hiccup: the harness crashed during MATH-12 when the Windows console (cp1252) could not print "2+√3"; it was restarted with PYTHONIOENCODING=utf-8 and MATH-12 was re-sat from the start. Earlier papers were not re-run.
• Marking quirks (scores left exactly as the scripts gave them):
  • MATH-12 M08: Muse answered 2+√3; the script expected the text 2+sqrt(3) and marked it wrong.
  • HOLE-16 P01: Muse answered -40 J; the key is -40; marked wrong on the unit.
  • WEAK-20 W07–W10 failed for Muse, Nemotron, Ornith, Qwen 3.8 27B and Qwen 3.6 with python rc=None; stdout='', so this is very likely the harness's Python runner, not the models. Siropia's sit (2 Nov) passed W07–W10.
  • Siropia MATH-12 M02, M07, M08, M11, M12 and MATH-HARD-12 MH03, MH10, MH11 hit the paper's token cap (800 / 1200) without an ANSWER line (thinking off, verbose working). That is 8 of her 15 lost marks; a larger max_tokens was not tested. M12 reached "01" (key 1) but wrote no ANSWER line.
  • Siropia's other fails: CLIFF-16 G01; HOLE-16 H03, H05; WEAK-20 W04, W06; CLIFF-24 T04, H02.
  • Siropia's sit: 166 calls, 0 errors. Strata has no /apply-template or /tokenize, so the harness context cap never applied; no call came near the server's 131072 context. VRAM in use was about 29 / 32 GB per card (nvidia-smi 29037 and 32047 MiB of 32768 after the speed tests; the expert cache fills free VRAM). Raw per-paper output: results/siropia-siona-strata-2026-11-02/.
  • Nemotron PHYS-12 P06 was marked FAIL with diagnosis "got 10 units expected 10".

Speed tuning (31 Oct – 1 Nov 2026)

Method for every table: llama.cpp llama-server, a fixed prompt of about 186–189 tokens (208 on the V100 box), 400 output tokens with ignore_eos, temp 0.2, 3 runs, timings read from the /completion response; values are the mean of the 3 runs. Draft runs vary a lot from run to run. "dflashN" = DFlash draft with n-max N; "tensor"/"layer" = split mode.

Muse Glimmer 30B, 2x Tesla P100 16GB (ctx 32768). Build A = existing llama.cpp (6fed9f6ff); Build F = qr0n/p100-llama-cpp community fork (dc5c644, upstream b10660 + P100 patches, sm_60).

Setup	gen tok/s	prompt tok/s
A layer, q8 KV, dflash16 (baseline)	5.16	81.5
A layer, no draft	13.15	92.3
A tensor, no draft	20.13	169.6
F layer, no draft	24.34	99.2
F tensor, f16 KV, no draft	40.24	212.8
F tensor, q8 KV, dflash4	51.41	180.8
F tensor, f16 KV, dflash2 (best)	54.82	179.2
F tensor, f16 KV, dflash6	31.77	183.6

About 10.6x generation and 2.2x prompt. With n-max 16 only about 18–35% of drafts were accepted.

Dell R730xd, 2x GTX 1070 8GB (cards at 100 W).

Model / setup	gen tok/s	prompt tok/s
Qwen 3.8 27B IQ4: layer, q8 KV, 32k (baseline)	9.13	118
Qwen 3.8 27B IQ4: tensor, q8 KV, 32k	15.1	163–170
Qwen 3.8 27B IQ4: tensor, q8 KV, 16k, MTP n4 (best)	25.3	140
SimpleSD 30B Q3_K_S: layer, fa off, 12k (baseline)	35.1	349
SimpleSD 30B Q3_K_S: tensor, fa on, f16 KV (best)	43.1	436

Qwen best is about 2.8x but halves context to 16k (f16 KV and MTP at 32k ran out of memory).

Muse Glimmer 30B, 2x Tesla V100 32GB (ctx 32768 unless noted; prompt processing about 400–490 tok/s throughout).

Setup	gen tok/s (mean)	draft accept
layer, q8 KV, dflash16 (baseline)	52.4	0.31
tensor, q8 KV, no draft	45.8	–
tensor, q8 KV, dflash16 (chosen)	71.8	0.37
tensor, f16 KV, dflash3	74.8	0.73
chosen setup at full ctx 131072 with vision (mmproj)	65.6	–

Tensor split with q8 KV and dflash16 was chosen as the most consistent fast setup with the smallest change; f16 KV with dflash3 was slightly faster at 32k.

Files

• <PAPER>/<PAPER>.jsonl: one item per line, with the fields id, category, prompt (for multi-turn items, a list of user turns), max_tokens, answer_key, accepted_forms, scoring_notes. Vision items also have an image field.
• <PAPER>/SPEC.md: the frozen specification. KEY.txt is included where it exists. run_*.py is the runner and mechanical scorer.
• CLIFF-24/verify_keys.py computes the CLIFF-24 keys independently. CLIFF-24/FROZEN.sha256 holds the hashes from freezing. The runner and verifier still match them; SPEC.md does not, because private details were removed from it before publication. Its text is otherwise unchanged.
• SHA256SUMS: hashes of every file as published.
• report.md: the write-up, with protocol, findings, the 31 Oct – 1 Nov auto-marked re-runs and speed tuning, the JUDGE-BREAK-32 scores and a small video review test.
• results/scores.csv and results/scores.json: per-model first-try scores. The mechanical score and the manual check are kept in separate columns.
• results/siropia-siona-strata-2026-11-02/: Siropia's raw nine-paper output (per-paper replies.jsonl, progress.txt, summary.txt, tally.json, saved artifacts under files/, plus calls.jsonl and stats.json). Local paths and the server address are replaced with <RUN_DIR> and <siona-host>; runner copies, SPEC/KEY copies and the WEAK-20 images are left out (the published originals are in each paper folder).

Notes

• Images withheld. The four WEAK-20 vision images (W15 to W18) are not included because their licence is unknown. Each item describes its image, and the keys are given.
• Model names. "Muse Glimmer 30B" is the open-weights model file that was run. "Siona" and "Sirin" are the names of Paul Buckingham's own local servers and deployments running it. "Siropia" is the local deployment name for Qwen3.8-Flash-Next IQ3_S on the Strata engine on Siona. These are local deployment names, not products. This dataset is not affiliated with or endorsed by any model vendor.
• Runners. Each runner talks to an OpenAI-compatible server (llama.cpp or vLLM). Set API and MODEL, and set ROOT to an output folder. The runners save every reply and artifact so that the manual check can read them.
• Some prompts contain negative constraints written for the author's own setting (for example "do not mention school"). They are left verbatim because the papers are frozen.
• JUDGE-BREAK-32 is a further paper. It is still private: its scores are reported in report.md, but its questions, answers and keys are withheld and are not part of this dataset.

Citation

@misc{axonis2026judgepapers,
  title        = {Axonis Pty Limited Judge Papers: small frozen test papers for local language models},
  author       = {{Axonis Pty Limited} and Buckingham, Paul},
  year         = {2026},
  howpublished = {Hugging Face dataset, axonis22/axonis-judge-papers},
  note         = {Retired papers; CC BY 4.0}
}