CoolFace
Datasetpublic

malaiwah/qwen3-5-tiny-cpu-repro-v1

Qwen3.5 tiny native random CPU fixture Complete randomly initialized, untrained Qwen3_5ForConditionalGeneration checkpoint. This is a pipeline/reproducibility fixture, not a useful language model, distillation, quantization, quality benchmark, or claim about the performance of Qwen3.8-27B. No upstream model weights or training data were used. No paid GPU/cloud compute. Architecture and lineage Architecture lineage: Qwen/Qwen3.8-27B at… See the full description on the dataset page: https://huggingface.co/datasets/malaiwah/qwen3-5-tiny-cpu-repro-v1.

sourceHugging Facemitupdated 17d agoView on Hugging Face
0likes249downloads
Dataset Card

Qwen3.5 tiny native random CPU fixture

Complete randomly initialized, untrained Qwen3_5ForConditionalGeneration checkpoint. This is a pipeline/reproducibility fixture, not a useful language model, distillation, quantization, quality benchmark, or claim about the performance of Qwen3.8-27B. No upstream model weights or training data were used. No paid GPU/cloud compute.

Architecture and lineage

Architecture lineage: Qwen/Qwen3.8-27B at 1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0. The native Transformers model type remains qwen3_5, with nested model.language_model, a real small model.visual, and an independent lm_head. Four dense-MLP text layers use explicit 3 linear / 1 full attention schedule: hidden 64, intermediate 128; gated full attention Q6/KV1 with head dimension 32; GatedDeltaNet QK1/V3 heads of dimension 16 and convolution width 4. Rotary dimension 8 uses nonempty mRoPE sections [2,1,1]. MTP is absent (zero tensors/layers). Vision: one real block, hidden 32/intermediate 64, four heads, patch2/temporal2, merge2, 16 learned positions and output64. All modality IDs are in vocabulary range. The tokenizer is genuine 256-byte BPE without merges plus 16 special/reserved tokens (total272); BOS/EOS processing and Unicode roundtrips are checked.

Text-only CPU fixture. Real small vision encoder and merger initialized, saved and reloaded; image/video execution and unequal-axis mRoPE are not tested. No processor is supplied. No processor is supplied; do not interpret the presence of vision weights as vision testing.

Measured CPU reproduction

  • —Generated parameters: 245,378, including 34,688 vision parameters.
  • —Stored semantic tensors: 77. Initialization seed: 20260907.
  • —Native FP32 storage exceptions observed by exact semantic reload: [].
  • —Native save/reload and QFS load compare values and dtypes, without float-cast masking.
  • —QFS loading report: observed=True, conversion errors visible=True; see full native-verification.json.
  • —Two fresh CPU captures, two PyTorch threads each, BF16, honest QFS lane other.
  • —Four synthetic64-token windows,252 shifted next-token positions; not benchmark data.
  • —Forced numerical replay metric: 0.0, KL {"max": 0.0, "mean": 0.0, "median": 0.0, "p95": 0.0, "p99": 0.0, "p99_9": 0.0}; top1 agreement 1.0.
  • —Native cached single-token and multi-token continuation runs include convolution and64-token chunk boundaries; exact within-schedule repetition is checked. Cached-vs-cold differences are measured, not asserted zero. Unequal-axis mRoPE is not exercised.
  • —Captured checkpoint pin: malaiwah/qwen3-5-tiny-random-bf16@49da461d1aaf0e3fa97b09f3490f0c7f6e0f9462. QFS source revision: 5cf8a81b8b701a2476fbed38619e62bd1052914e (source hashes accompany evidence).

Evidence and limitations

Evidence bundle contains first/ and repeat/ as independently sealed QFS root datasets, comparison/, strict verification outputs, actual per-process runtime/resource records, command logs, the sealed panel and its tokenizer binding. The repository root is an artifact receipt bundle, not a single root-format dataset. Reproduction confirmation retains QFS's hash-proof semantics; --force-compute also runs numerical replay and verifies252 positive float64 zero KL values. This is same-machine, same-stack repeatability, not cross-hardware equivalence or model quality evidence. See evidence-summary.json, build-manifest.json, native-verification.json, qfs-source.json, and hub-checkpoint-verification.json.

Reproduce locally

Use Python3.12 and an isolated environment with pip install -r requirements-cpu.txt. No FLA, causal-conv1d, CUDA packages, remote model Python, or GPU is needed. With the recorded QFS checkout available at $QFS:

bash
python build_fixture.py --out /tmp/qwen35-rebuilt
python verify_native.py --model-dir /tmp/qwen35-rebuilt --qfs-root "$QFS" --out /tmp/qwen35-native.json
python build_panel.py --model-dir /tmp/qwen35-rebuilt --model-repository malaiwah/qwen3-5-tiny-random-bf16 --model-revision 49da461d1aaf0e3fa97b09f3490f0c7f6e0f9462 --qfs-root "$QFS" --out /tmp/qwen35-panel
python verify_cpu.py --model-dir /tmp/qwen35-rebuilt --model-repository malaiwah/qwen3-5-tiny-random-bf16 --model-revision 49da461d1aaf0e3fa97b09f3490f0c7f6e0f9462 --panel-dir /tmp/qwen35-panel --qfs-root "$QFS" --out /tmp/qwen35-repro

Outputs must not exist. release.py --stage-only --out NEW_DIR --qfs-root "$QFS" is the publisher workflow: it requires write authorization for the two fixed repository IDs, refuses repository collisions, and stages privately. Explicit --publish with the same output directory only releases a successful unchanged stage and verifies anonymous downloads. It intentionally cannot be used to overwrite these published repositories.