aditiPatodiya/quantcheck-audit

quantcheck census of quantized LLM artifacts

This dataset is the execution census behind the paper Broken on Arrival: Non-Functional Quantized LLM Artifacts in Public Model Registries and How to Catch Them (Aditi Patodiya, 2026; arXiv:2609.05881, DOI 10.48550/arXiv.2609.05881).

We ran 327 code-capable GGUF artifacts (the llama.cpp file format) on a 15-task HumanEval+ screen and kept every generation. The artifacts are all 305 eligible tags of 15 model lines in the official Ollama library and 22 popular community files from Hugging Face. Each row of audits is one artifact, with its scores, its class, its file hash and, for the seven suspects, the paper's attribution. Each row of generations is one model answer with its pass or fail verdict.

The tables are built from release v3.4 of the quantcheck GitHub repository. The repository also holds the rebuild, calibration and tensor-comparison runs, the screening tool, and the scripts that recompute the paper's numbers; its PROVENANCE.md marks what rests on inputs that are not released.

Findings

• Seven non-functional official tags. Seven of the 305 official tags (2.3%) are suspects on the screen and are confirmed on all 164 HumanEval+ tasks. Six solve none and llama3.2:1b-instruct-q3_K_S solves 6. They are four qwen2.5-coder:3b-instruct tags (q2_K, q3_K_S, q3_K_M, q3_K_L), phi3.5:3.8b-mini-instruct-q2_K and two llama3.2:1b-instruct tags (q2_K, q3_K_S). Three of the 35 model variants are affected (8.6%, Wilson 95% interval 3.0% to 22.4%). Tags of one variant are not independent, so the variant rate is the better summary; per tag, the interval is 1.1% to 4.7%.
• No file defect among the five rebuilt. We rebuilt five of the seven from the original checkpoints with llama.cpp and no importance matrix. The four Qwen2.5-Coder-3B files match our rebuilds in all 434 tensors. The phi3.5 file matches our rebuild in all 197 tensors once the rebuild's 32 attn_qkv tensors are set to Q2_K, where llama.cpp v0.7.0 uses Q3_K. These five registry files are exact, uncalibrated quantizations of the released weights. The license-gated llama3.2-1B checkpoint was not rebuilt, so its two tags are not attributed; calibrated third-party files of the same model and levels also fail, at 1 and 4 of 15.
• Calibration partly rescues four of the five. Adding only an importance matrix to the same weights, toolchain and tensor mix raises Qwen q2_K from 0 to between 106 and 113 of 164 tasks, q3_K_M to between 76 and 89 and q3_K_L to between 80 and 86. Qwen q3_K_S stays weak under every calibration (13 to 30). For phi3.5, the tensor mix of llama.cpp v0.7.0 lifts the uncalibrated build to 21 and an importance matrix to 37.
• Community files. 20 of the 22 Hugging Face files score at least 9 of 15 in the census. The other two, TheBloke's deepseek-coder-6.7B files at Q3_K_M and Q4_K_M, score 5 and 6. The census supplied no chat template for community files, and these two files carry none, so the runtime fell back to a generic ChatML format that differs from the model's documented instruction format: 28 of their 30 answers run on into invented chat turns until the token cap. With the deepseek-coder instruction template supplied, the same files score 13 and 15 of 15 on an RTX 4090 with the same Ollama release (98 and 101 of 164) and 13 and 15 of 15 on Apple Metal. Ollama's own conversions of the model score 14 at both levels. They were not rebuilt; the paper records them as template-dependent failures, not file defects.
• Seven is a lower bound. All 12 deepseek-coder-1.3B tags, fp16 included, score 1 to 6 of 15, so a group baseline cannot judge them, and the group is not attributed. Full runs of flagged tags that are not suspects find eight more files that solve at most 16 of 164: seven in this group and gemma3:270m-it-q8_0.
• The screen's misses were audited. 66 tags that passed the screen were run on 49 other HumanEval+ tasks and solved 8 to 41 of them. A non-functional file solves at most 16 of the 164 tasks, and for each of the 66 a tested file solves at least 17 distinct tasks: 50 on the audit tasks alone, 14 with their screen tasks on the same file and backend, and 2 in full reruns on the census files (65 and 44 of 164). One tag, gemma3:4b-it-q8_0, was re-published before the audit and was tested on its newer file. For the 223 tags outside the tuning set and the first-version sample that passed the screen as first scored, the 56 seeded draws give an exact stratified 95% upper bound of 22 non-functional census files (9.9%), counting the re-published tag's August file as failed.

Read this before using the columns

A low score says that an artifact did not work in this setup, not why.

Column	What it is	What it is not
stage1_flag	Screen score below 9 of 15. A trigger for a closer look (39 of 305 official tags).	A verdict. 31 of the 39 flagged tags are models of at most 4B parameters, and 23 of the 39 are healthy against their own baseline.
stage2_class	The tag's class against the median score of its model variant at q6_K, q8_0 or fp16 (group_baseline).	A cause. "low baseline" means the variant scores low even at high precision.
confirmed_nonfunctional	A suspect that also solves under 10% of the 164 HumanEval+ tasks. True for exactly seven tags.	A statement about the file's integrity, or proof that a false row works: eight flagged non-suspects solve at most 16 of 164 (see full164_score and gate_route).
attribution	The verdict column of the paper's Table IV for the seven suspects: based on a rebuild from source for five, not attributed for the llama3.2-1B pair.	A judgement on any artifact where it is empty.

The raw census files in the quantcheck repository also carry a field named verdict, written by the first-version screen. It is the stage-one flag on the original scores, before the HumanEval/15 correction (50 official tags against 39 here), and it is not included. Version 1 of the preprint described the four Qwen2.5-Coder-3B tags and the phi3.5 tag as defective conversions. Recipe-matched rebuilds showed that they are exact uncalibrated quantizations, and the revised paper (arXiv v2) withdraws that diagnosis.

Configs

audits: one row per artifact (327 rows)

Column	Type	Meaning
tag	string	Ollama tag, or hf:<uploader>/<repository>/<file> for a Hugging Face file
channel	string	ollama-official (305) or hf-community (22)
model_line	string	Ollama model name before the colon; for HF files, the repository name without -GGUF
model_variant	string	Official tags: the tag without its quantization suffix, the stage-two group (35 variants). HF files: the repository id
quantization	string	Quantization level as written in the tag or file name
params_b, file_size_gb	float	Parameter count in billions and file size in GB, from the census inventory
url	string	Library page of the tag or of its model's tag list, or the download URL, as in the census inventory
file_sha256	string	SHA-256 of the GGUF file: the registry blob the census ran for official tags; for HF files, the downloaded file the census imported into Ollama, which loaded a re-written copy. Extracted from the census server log (not released) into rerun-2026-10/serial/census_digests.json
run_started_utc	string	Start of the artifact's census run (pull or download, then the screen)
screen_score	int	HumanEval+ tasks passed on the 15-task screen (serial scoring, below)
screen_passed_tasks	list	Task ids of the passed screen tasks
stage1_flag	bool	screen_score below 9
group_baseline	float	Median screen score of the variant's q6_K, q8_0 and fp16 tags; empty when the variant has none within the 8 GB frame, and for HF files
stage2_class	string	healthy (at least 60% of the baseline), suspect (at most 2 of 15 with a baseline of at least 8), low baseline (below 60% with a baseline below 8), degraded (any other score below 60%) or no baseline. Official tags only
full164_score	int	HumanEval+ tasks passed out of 164 on the census file, where a full run exists (31 artifacts)
full164_run	string	census, August 2026, October 2026 re-download, same file as census or miss-audit closure, October 2026, census file
confirmed_nonfunctional	bool	Suspect and full164_score under 10% of 164
metal_screen_score	int	Screen score on Apple Metal: the five rebuilt suspects, run as registry tags, and the two TheBloke files, run with the deepseek-coder instruction template supplied (the census supplied none, and the runtime used its generic ChatML default; the CUDA runs with that template are in the repository under rerun-2026-10/results-tpl/)
attribution	string	collapse, rescued by calibration, collapse, rescue rule not met or not attributed (no rebuild). The seven suspects only
best_calibrated_full164	int	Best calibrated build at the same level on all 164 tasks, from any source, as in the paper's Table IV. For q3_K_M this is Qwen's own GGUF file labeled Instruct-AWQ (125), which also stores a separate output head; our importance-matrix builds reach 76 to 89. The phi3.5 build uses the tensor mix of llama.cpp v0.7.0
gate_route	string	Where the paper's proposed acceptance gate sends the tag: screen passed, suspect: attribution, flagged, full run 17 or more of 164: publish, flagged, full run 16 or fewer of 164: withhold or flagged, no full run: pending. The gate's perplexity pre-check was not run on the census and is not reflected. Official tags only
miss_audit_sample	string	For the 66 audited tags: first version, August 2026 or seeded draw, October 2026
miss_audit_heldout49	int	Tasks passed out of the 49 held-out tasks in the miss audit. For gemma3:4b-it-q8_0 the audit ran on the re-published October file, not the census file; for the 10 first-version tags the audited file's digest was not recorded
truncation_rate	float	Share of the 15 screen generations that hit the 768-token cap
prompt_echo_mean	float	Mean share of trimmed prompt lines longer than 12 characters that appear verbatim in the output
repetition6_mean	float	Mean repetition score: six times the count of the most frequent word 6-gram divided by the word count (0 below 12 words). It can exceed 1
def_header_rate	float	Share of screen outputs with a line matching the one-line header pattern def name(...):. Headers with a return annotation (def f(x) -> int:) do not match, so working tags often score about 0.33

generations: one row per model answer (9,333 rows)

The screen answers of all 327 artifacts (4,905 rows) and the answers of the 27 census full runs on all 164 tasks (4,428 rows). The October 2026 reruns are in the quantcheck repository, not here.

Column	Type	Meaning
tag, channel	string	As in audits
run	string	screen (15 tasks) or full164 (164 tasks)
task_id	string	HumanEval+ task id
in_screen_suite	bool	Whether the task is one of the 15 screen tasks
passed	bool	Serial EvalPlus verdict: both the base and the strengthened tests pass
output	string	The model's full response, unedited except for one answer of TheBloke's Q4_K_M file, where an email address reproduced from training data is replaced by <EMAIL_REDACTED>. Outputs may contain URLs and user names that the models reproduced or invented
hit_token_cap	bool	Whether generation stopped at the 768-token cap
generated_tokens	int	Tokens generated
repetition6, prompt_echo	float	Per-answer repetition score and prompt echo, as in audits
def_header	bool	Whether the answer has a line matching the header pattern, as in audits
transcript_file	string	Source transcript in the quantcheck GitHub repository, release v3.4

The five per-answer statistics are recorded for screen answers only and are empty for full runs. hit_token_cap and generated_tokens are also empty for 10 screen answers for which Ollama returned no completion record.

How the data were produced

• Census. The inventory was generated on 10 August 2026 and the census ran on 12 August 2026 on one NVIDIA RTX 4090 with Ollama 0.32.6. A tag was eligible if it was an instruct, code-capable build with an explicit quantization suffix and a file of at most 8 GB: 305 tags of 35 variants, 291 quantized and 14 fp16. The 22 community files are Q2_K to Q4_K_M files from 12 Hugging Face repositories. The census supplied no chat template for them, so Ollama used each file's own chat template, or the runtime's generic ChatML default where a file has none.
• Decoding. Ollama chat endpoint, temperature 0, seed 42, context 2,048 tokens, at most 768 new tokens.
• Screen. The 15 tasks are the first 15, in lexicographic order of task id, of the 39 HumanEval+ tasks that all 23 healthy configurations of a separate Qwen2.5-Coder tuning study solve. On that tuning set every healthy configuration passes 15 of 15 and the one known failure passes none. Screening costs about 30 s of generation per tag.
• Scoring. EvalPlus 0.3.1 plus-tests. The census scorer marked HumanEval/15 as failed for every artifact because its strengthened tests can exceed the time limit when scoring runs in parallel. Every failed task was re-evaluated serially, one evaluation at a time with default limits, together with a 3% sample of passes, none of which changed. HumanEval/15 then passes for 275 of the 305 official tags. All scores here are the serial scores.

Not included

In the quantcheck GitHub repository (release v3.4), not in this dataset:

• The October 2026 reruns (rebuilds, calibration, generation settings, miss-audit runs, and the template test of the two TheBloke files) under rerun-2026-10/, including the file hashes and per-tensor comparisons of the rebuilt, calibrated and third-party files.
• The first-version miss-audit runs under results-sensitivity/, the Metal replication under results-mac/ and results-metalrep/, and the control runs of the two TheBloke files under results-native/.
• Seven Hugging Face files that version 1 used as referees, run on the census machine on the same day. They are not part of the census and are in the repository's results/ files with channel hf-adjudication.
• The per-task results of the tuning study's 24 configurations.

Not released anywhere: the rebuilt, calibrated and third-party GGUF files themselves, which the repository identifies by their hashes, and the tuning study's generations.

License

The data are released under CC BY 4.0, as in the quantcheck repository; its code is under the MIT License. The task prompts come from HumanEval (MIT License) as extended by EvalPlus (Apache License 2.0), and many outputs repeat them. Model outputs may also be subject to the terms of the model that produced them.

Citation

@article{patodiya2026brokenonarrival,
  title   = {Broken on Arrival: Non-Functional Quantized {LLM} Artifacts in
             Public Model Registries and How to Catch Them},
  author  = {Patodiya, Aditi},
  journal = {arXiv preprint arXiv:2609.05881},
  year    = {2026},
  doi     = {10.48550/arXiv.2609.05881},
  url     = {https://arxiv.org/abs/2609.05881}
}

Loading

from datasets import load_dataset

audits = load_dataset("aditiPatodiya/quantcheck-audit", "audits", split="train")
generations = load_dataset("aditiPatodiya/quantcheck-audit", "generations", split="train")