The dataset viewer is not available because its heuristics could not detect any supported data files. You can try uploading some data files, or configuring the data files location manually.
Does 3-bit quantization break Arete-27B's value alignment?
Arete-27B exists to change one behaviour. On the publisher's own evaluation it cut answers containing propaganda from 22% to 0.9%. Quantization to roughly three bits can damage a narrow behaviour like that while leaving general benchmarks intact, and 47 derivative repositories had shipped quantized builds of this model without anyone measuring it.
This dataset is that measurement: three builds of the same model, the publisher's own rubric and grading code, one judge model across every row.
Results
402 prompts, one judge for all rows, 0 of 1206 verdicts failed to parse.
| Build | Size | bits per weight | Fact mean (0-5) | Any propaganda |
|---|---|---|---|---|
| BF16 original | 54.71 GB | 16.000 | 2.440 | 11.44% |
| IQ2_M with SSM and attention protected | 10.88 GB | 3.181 | 2.092 | 11.19% |
| bartowski IQ2_M, stock layout | 10.28 GB | 2.880 | 2.139 | 10.95% |
The value alignment survives
The propaganda rate does not move. Against the BF16 row the differences are 0.25 and 0.49 percentage points, with a standard error of about 2.2, so z is −0.1 and −0.2. At three bits per weight the property the model was built for is intact as far as this metric can see.
Factual recall does degrade
The fact mean falls from 2.440 to 2.092, a loss of 0.348 on a 0 to 5 scale against a standard error of the difference near 0.14, so z is about −2.5. That is roughly 12% of the factual recall for an 80% reduction in file size.
Protecting the recurrent tensors bought nothing
The 10.88 GB build holds the ssm_alpha and ssm_beta tensors of all 48 Gated DeltaNet layers
at Q8_0 and the attention projections of all 16 full attention layers at Q5_K, which costs 0.30
more bits per weight than the stock IQ2_M. Against that stock build it measures 0.047 lower on
fact mean and 0.24 percentage points higher on propaganda, both inside the noise. The damage from
low-bitrate quantization is not concentrated in the tensors a reader of this architecture would
suspect first, and the standard layout is already efficient here.
Because that build has no measured advantage over a file 0.6 GB smaller, it is not published as a model. There is no reason to prefer it.
The headline number depends on the judge, not only on the model
The Arete-27B card reports 0.9% propaganda for the BF16 weights, graded by x-ai/grok-4.7. The
same weights, the same decoding protocol, the same prompts and the same rubric, graded by
anthropic/claude-haiku-5.5, give 11.44%.
This is not a claim that either figure is wrong. It is a measurement of how far an LLM-graded metric can move when only the grader changes, and the distance here is a factor of thirteen. A reader comparing propaganda rates across model cards should check which judge produced each number before treating the comparison as meaningful.
What is in here
| Path | Contents |
|---|---|
data/items-sample.jsonl |
the 402 prompts used, drawn from Propaganda_Eval |
data/completions-base.jsonl |
BF16 answers |
data/completions-mine.jsonl |
10.88 GB build answers |
data/completions-ref.jsonl |
bartowski IQ2_M answers |
data/grades-*.jsonl |
one judge verdict per prompt, with the facts it marked stated, missed and opposed |
data/scores.json |
score.py output for all three rows |
method/REPRODUCE.md |
every command, every parameter, every deviation |
method/grade-patched.py |
the single-line change to the official grader, and why it was needed |
method/run_eval.py |
the answer generation client |
Limitations
402 of 803 prompts, chosen as every second prompt by sorted id, because of a judge budget limit rather than any methodological reason. All nine categories are present at 49 to 53 percent of their full counts.
One judge model, not an ensemble, and not the judge the publisher used. The row-to-row comparison is internally consistent; the absolute values are not comparable to the published figures, which is itself one of the findings above.
Three of 803 answers in the full generation hit the per-slot context bound and are truncated. The bound applied equally to every row.
The multi-token prediction head was excluded from all three GGUF builds, so none of them serves as a speculative decoding draft.
Credits
Nous Research for Arete-27B and for releasing the evaluation, its rubric and its grading code as open artifacts. Without the grading code this check would have measured something else and been worth nothing.
bartowski for the importance matrix and its calibration corpus, reused unchanged, and for the stock IQ2_M build that serves as the third row.
Vladimir Beskorovainyi, Enterprise AI Architect & CTO, author of the Amazon bestseller "Enterprise AI: The KATA Way", founder of Besk Tech (besk.tech)
- Downloads last month
- 64