malaiwah/glm53-fixture-0.1B-fidelity-quant-int4-v1

GLM-5.3-Flash-0.1B fixture — candidate fidelity dataset, toy RTN-int4 routed experts (hidden form)

The numbers in this dataset are meaningless as quantization quality. The weights are random (inference-optimization/GLM-5.3-Flash-0.1B-A0.1B is an architectural fixture), and the quantizer is deliberately crude. This exists so that step 3 of the three-step fidelity architecture has two real datasets to compare, and so that anyone can see what a candidate capture looks like next to its root.

This is B. Its root is `malaiwah/glm53-fixture-0.1B-fidelity-root-v1`.

The candidate

bin/toy_quantize.py, exactly:

w_hat[i, j] = round(w[i, j] / s[i]) * s[i],   s[i] = max_j |w[i, j]| / (2^(bits-1) - 1)

applied to every tensor whose name contains .mlp.experts. — the 48 routed-expert gate_proj / up_proj / down_proj matrices, 1,572,864 elements — and to nothing else. 175 other tensors are copied through byte-for-byte, including lm_head, embed_tokens, the shared experts, every norm and every attention projection. There is no calibration, no error feedback, no group structure and no outlier handling. The result is dequantized back to bf16 and written as an ordinary safetensors checkpoint.

scope.assignments says exactly this: moe.experts is quantized: int4 @ 4, every other class is native: bf16 @ 16, and head_policy is native. Because the head is untouched and shipped identically on both sides, hidden replay through one shared head is legitimate here (HEAD-1c does not apply: the captures are not bitwise equal).

The measurement

The sibling int8 candidate scores 0.000154605 nats at 99.33% top-1. The ordering is the point; the magnitudes describe random weights.

python3 bin/fidelity_dataset.py compare \
  --reference hf://malaiwah/glm53-fixture-0.1B-fidelity-root-v1 \
  --candidate hf://malaiwah/glm53-fixture-0.1B-fidelity-quant-int4-v1 \
  --out cmp

Specification: <https://github.com/malaiwah/glm53-flash-fidelity-suite/blob/main/docs/FIDELITY-DATASET-SPEC.md>

MIT licensed.