Dontbeafed69/PonderScope-Qwen3.5-4B-Quantization-Reasoning

PonderScope — Qwen3.5-4B BF16 versus MLX Q4 reasoning evidence

Dataset summary and research objective

A public research-evidence compilation of existing controlled experiments comparing native BF16 and MLX affine Q4 deployments of the same pinned Qwen3.5-4B source. It includes viewer-compatible per-draw and paired Parquet tables, verified source-seal records, raw generation traces, procedural task populations, exact specifications, the frozen replication population lock, corrected derived comparisons, and publication checksums.

The question is whether a controlled change of executable weight representation changes measured reasoning behavior under a fixed runtime, prompt policy, decoding condition, and finite observation horizon. This is not a general intelligence benchmark, an equivalence test, or a peer-reviewed publication. No research executions were regenerated for this release.

Model source, deployment definitions, and conversion

Base model developer: Qwen / Alibaba Cloud. Qwen/Qwen3.5-4B, immutable revision 851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a.

Arm	Weight variant	Deployment	Definition
BF16	wvar-9b6f9b40b789	dep-e1c20569173a	Native upstream BF16 text execution, no quantization
Q4	wvar-b7a30758909a	dep-4f8e7958df4f	Derived MLX affine 4-bit/group-64 text weights

Both share source artifact src-6209ef70e318. MLX-LM 0.32.0 converted the source with BF16 conversion dtype and existing affine quantization. Actual detection: 248 quantized linear modules and 4.502633743 effective bits per weight. Unquantized components remain at their supported precision. No training or fine-tuning; no new quantization algorithm. Both deployments execute text only. The associated converted model contains the weights and exact conversion provenance.

Hardware/runtime: Apple M3 Pro, 19,327,352,832 bytes unified memory, MLX Metal, MLX-LM 0.32.0 / MLX 0.32.3, Python 3.12.12, Darwin 27.2.0. Captured package versions and scientifically relevant machine metadata remain in each bundle.

Task families, difficulty, and populations

tasks-v1 contains original, procedurally generated tasks with deterministic reference answers: arithmetic expressions (arith), graph paths (path), ordering constraints (order), logic constraints (logic), and state-machine simulation (sm). Structural identities and rendered presentation identities are distinct; tasks and prompts are included in the evidence archives.

Study	Population	Structural tasks	Seeds	Draws per arm	Per-draw rows	Matched pairs
Phase 1.4 calibration	calibration split, task seed 0	10; 2/family	0, 1, 2	30	60	30
Phase 1.6 independent replication	held-out test split, task seed 1729	30; 6/family	0, 1	60	120	60

The replication population was locked before inspecting outcomes and has no structural-task overlap with calibration. Calibration has 5 easy / 5 medium tasks; replication has 10 easy / 10 medium / 10 hard tasks. Difficulty is procedural task metadata, not a standardized human rating. 180 model executions across 40 distinct structural tasks and two deployment configurations. Pilot runs, other model sizes, technical-repeat diagnostics, and subsequent research are excluded from these tables and totals.

Data generation, scoring, and sampling policy

Prompt policy pp-v1, Qwen native thinking enabled, minimal capture lane, 2048 total generated-token horizon. Sampled decoding condition cond-f5e762d3078f: temperature 1.0, top-p 0.95, top-k 20, min-p 0.0, presence penalty 1.5, repetition penalty 1.0, frequency penalty 0.0. Presence penalties apply to the full generated history, excluding prompt tokens, through PonderScope's explicit processor. They are not MLX-LM's default rolling-window presence semantics. Exact specifications and processor metadata are preserved in each archive. One execution per task/seed/arm, repeat index 0.

Reference answers come from the task generator. Natural final-channel output is parsed and normalized with the task-family scorer; normalized scorable answers are checked against ground truth. Corrected outcome classification is recomputed from immutable record fields using the reviewed Phase 1.7 source. Historical stored classifications remain separately available. Generated reasoning is model output, not a human-authored explanation or certified solution.

Outcome and censoring definitions

• native_reasoning_closure: the model emitted its native thinking-close token. It is distinct from generation EOS or final-answer correctness.
• generation_terminated_by_eos: a natural generation EOS was observed.
• success_at_budget: a correct scorable answer was produced inside the horizon. False includes censored/unparseable observations and does not imply an incorrect final answer was observed.
• correct / incorrect: observed, normalized answers match / fail ground truth.
• unparseable: a raw final-channel answer is present but cannot be normalized; it is not classified as a wrong answer.
• censored: no natural closure was observed before the token cap.
• budget_exhausted_after_closure: generation was capped after reasoning closed; this is a separate flag from within-reasoning censoring.
• terminated_no_closure and error: distinct nonclosure termination/failure categories when applicable; no such failures occurred in the four main runs.

A right-censored reasoning trace does not establish an incorrect final answer. The 2048-token horizon is a measurement limit, not the model's natural stopping point. Reasoning-closure event time is reasoning_tokens + 1, including the native close; EOS time is the independent generation-terminal token. total_tokens includes delimiters and need not equal reasoning plus answer counts. Loop diagnostics are descriptive; repeated phrasing is not proven looping.

Statistical unit and pairing

One table row is a stochastic draw for one deployment. Pair by study, structural task, presentation, decoding condition and sample seed, excluding weight/deployment identity. The paired table contains 30 calibration and 60 replication rows. Technical repeats are not independent draws; none are included here. Every matched draw is unique and nonambiguous.

Uncertainty uses structural tasks as independent bootstrap clusters, keeping seeds paired within tasks (2,000 resamples, bootstrap seed 0). One execution per stochastic draw means within-deployment repeat noise is not separately estimable. Calibration and replication populations have different difficulty mixes and are not pooled into one unqualified estimate.

Primary results

All differences below are Q4 minus BF16.

Endpoint	Calibration estimate [95% CI]	Independent replication estimate [95% CI]
Native reasoning closure rate	+0.033 [0.000, +0.100]	+0.067 [-0.033, +0.183]
Restricted mean reasoning-closure tokens, horizon 2048	-117.3 [-307.9, +63.5]	-48.1 [-171.9, +62.3]
Success at budget	0.000 [-0.133, +0.133]	+0.050 [-0.017, +0.133]

Restricted-mean values above are the bootstrap means reported historically; empirical point differences are -117.5 and -48.8. Both values and full-precision intervals are recorded in metadata/primary-results.json. Native closure counts are BF16 20/30 versus Q4 21/30 in calibration and BF16 27/60 versus Q4 31/60 in replication. Success counts are 19/30 in each calibration arm and 26/60 versus 29/60 in replication. All primary intervals include zero.

Performance (arithmetic mean per draw)	Calibration BF16 / Q4	Replication BF16 / Q4
End-to-end generation tokens/sec	13.4 / 41.7	12.8 / 37.0
Decode-only tokens/sec	13.5 / 42.1	12.9 / 37.3
Generation wall seconds	112.6 / 33.0	124.1 / 41.9

The historical 12.8/37.0 headline includes first-token latency; decode-only rates are separately exposed. Short model-load audits recorded MLX allocator peaks around 8.53 GB BF16 versus 2.56 GB Q4 after generation; they are separate audits, not full-study matched peak-memory measurements.

Q4 delivered a large speed and memory advantage on the measured Mac/MLX setup, while the study did not establish a statistically distinguishable change in the primary reasoning outcomes. Neither study establishes BF16/Q4 accuracy equivalence, reasoning improvement, universal generalization, or superiority to other conversions. The repeated-4gram observation is exploratory and reverses sign between populations; it is not a primary loop-reduction conclusion.

Files and practical use

• data/calibration.parquet: 60 per-draw rows.
• data/replication.parquet: 120 per-draw rows.
• data/paired-calibration.parquet: 30 matched-pair rows.
• data/paired-replication.parquet: 60 matched-pair rows.
• data/paired-outcomes.parquet: 90 paired rows with an explicit study column; this is a convenience concatenation, not a pooled estimate.
• metadata/: comparison/paired results, normalized-direction primary results, schema/data dictionary, frozen population lock, source and public verification reports, model-load audits, export provenance and privacy records.
• evidence/{calibration,replication}-{bf16,q4}.tar.gz: four archives containing raw traces, tasks, environment, source seal, sanitized manifest, derivative integrity seal, exact spec, historical/current analyses, and redaction ledger.
• SHA256SUMS: SHA-256 for every published release file except itself.

All 180 main-study executions have raw evidence coverage. Token sequences, generation text, prompts, ground truth, termination observations, timing, metrics and loop diagnostics are preserved. Trace text is stored once in its archive; the tables expose answer fields and exact archive/member/line references.

After installing datasets, use the Hub's viewer-compatible Parquet configuration:

from datasets import load_dataset

repo = "Dontbeafed69/PonderScope-Qwen3.5-4B-Quantization-Reasoning"
calibration = load_dataset(repo, "draws", split="calibration")
replication = load_dataset(repo, "draws", split="replication")
pairs = load_dataset(repo, "paired_outcomes", split="replication")
assert len(calibration) == 60 and len(replication) == 120 and len(pairs) == 60

Per-draw columns distinguish study/population, structural/presentation identity, source revision, variant/deployment identity, condition/seed, ground truth, raw/normalized answer, historical/corrected outcome, closure/EOS/censor flags, tokens, timing, both throughput definitions, repetition/loop diagnostics and raw references. See metadata/table-schemas.json and metadata/DATA_DICTIONARY.md. Pair transitions use both, neither, only_a (BF16 only), only_b (Q4 only). neither in success-at-budget is not “both answered incorrectly”. Source comparison JSON uses A(BF16) minus B(Q4); normalized primary results reverse that sign explicitly for Q4-minus-BF16 presentation.

Privacy, provenance, and evidence integrity

The original immutable evidence remains private and unchanged. Original seals verified locally before export. Public bundles are sanitized derivatives: local home-directory prefixes were replaced with /REDACTED_HOME in metadata and deployment-path fields. No scientifically relevant generation field was changed. Each redactions.json identifies file/JSON locations and original and published file hashes, hashing original string values rather than exposing them. Local runtime/hardware metadata needed to interpret the condition is retained.

source-evidence-seal.json is the byte-identical historical seal. It describes original bytes, not the sanitized manifest/trace bytes. evidence.json is a clearly labeled release-day integrity seal over the public derivative, compatible with the existing verifier. These hashes provide integrity, not a cryptographic signature or independent attestation. Public source authenticity depends on the recorded local original verification and retained private originals; readers cannot reconstruct redacted paths from public hashes. Published checksum values must not be mistaken for the original immutable seal values.

Known corrections and versioned reanalyses

Phase 1.4 corrected closure time to the native close-token index; Phase 1.5 corrected paired restricted-mean estimation, independent EOS event timing and evidence gating; Phase 1.6 clarified condition-aware draw identity and paired population alignment. Phase 1.7 corrected raw-but-unscorable answers from incorrect to unparseable and separated capped-after-closure observations. One replication draw in each arm becomes unparseable; primary results are unchanged. analysis.phase1_6.json is retained where available. Historical stored outcomes in raw traces remain unchanged; table outcomes are recomputed. The early failed-seal smoke and all subsequent prefix-probe research are excluded.

Reproducibility instructions

Download the dataset with the supported CLI:

hf download Dontbeafed69/PonderScope-Qwen3.5-4B-Quantization-Reasoning \
  --repo-type dataset --local-dir ponder-evidence
cd ponder-evidence
shasum -a 256 -c SHA256SUMS
mkdir unpacked
for archive in evidence/*.tar.gz; do tar -xzf "$archive" -C unpacked; done

For offline verification/reanalysis, use the reviewed PonderScope source in a separate checkout. Install with uv sync --frozen --extra dev; run uv run ponderscope verify --run <unpacked-run-directory> --json for each archive. That verifies the public derivative against its new integrity seal. Then use ponderscope analyze or ponderscope compare --a <bf16-run> --b <q4-run> --mode sampled --require-complete; no model inference is required. Preserve historical analyses before writing new reports. The archive's source seal remains separately identified. The export utility in the source repository refuses unverified inputs, incomplete pairs, unexpected repeats and scientific-content redactions. Offline reanalysis is reproducible; historical stochastic token sequences are not promised to be regenerated byte-for-byte by rerunning a model.

Uncertainty and limitations

One model/source revision, one Apple Silicon Mac/runtime, 10 and 30 task clusters, small synthetic tasks, correlated seeds within tasks, finite horizon and extensive family-specific censoring. logic has zero native closures in both replication arms. Lack of a detected difference is not evidence of equivalence. No trained controller, policy-transfer claim, causal explanation of family differences, or broad Qwen-family conclusion is established. Outcomes outside the horizon are unobserved. Generated text may contain errors and should not be treated as advice.

License and citation guidance

MIT applies to PonderScope's original research compilation, generated task material, instrumentation-derived records, and documentation, consistent with the MIT task-generator/scoring code. The inputs are original procedural tasks, not a third-party benchmark corpus. Generated traces are outputs of the attributed Qwen model. Dataset licensing does not relicense Qwen's pretrained weights or grant rights in independently protected third-party content. Qwen's exact upstream weights license is Apache-2.0; weights are distributed only in the associated model repository with the appropriate notices. See LICENSE and NOTICE.

Cite this as a versioned public research resource, not a paper: “PonderScope contributors (2026), PonderScope Qwen3.5-4B Quantization Reasoning Evidence, Hugging Face dataset, released 2026-10-12,” followed by the dataset URL and exact Hub revision used. Credit Qwen for the underlying model. Report the study split, source revision, deployment, decoding condition and analysis version in reuse.

Source GitHub repository · Associated MLX Q4 model · Research limitations.