File size: 2,872 Bytes
ffdd9aa
 
f3a14ee
ffdd9aa
 
 
 
 
 
 
 
 
 
 
 
 
 
f3a14ee
 
 
ffdd9aa
 
 
 
 
 
 
 
 
 
 
 
f3a14ee
 
ffdd9aa
 
 
 
 
 
 
 
 
 
 
 
 
 
f3a14ee
ffdd9aa
 
 
 
 
 
f3a14ee
ffdd9aa
 
 
 
 
 
 
 
 
 
f3a14ee
ffdd9aa
 
 
f3a14ee
 
 
 
 
 
 
ffdd9aa
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
---
pretty_name: Embedding Quantization Reproducibility Bundle
license: apache-2.0
tags:
- embeddings
- quantization
- mlx
- zerogpu
- reproducibility
---

# Embedding quantization reproducibility bundle

This dataset contains the frozen inputs, vector artifacts, metrics, calibration
evidence, and scripts behind the companion engineering post on MLX Q/oQ/oQe
embedding quantization and CUDA-native controls.

See [`RESULTS.md`](./RESULTS.md) for the compact result tables and evidence
boundaries.

The model weights are not duplicated here. `models.lock.json` pins all 30 public
model repositories to exact Hub commits and records their canonical upstream
BF16 revisions. Quantized artifacts branch directly from their family's BF16
checkpoint; no lossy conversion was used as another quantization's source.

## Verify the published evidence

```bash
hf download TiGa-RCE/embedding-quant-repro-2026-07-28 \
  --repo-type dataset --local-dir embedding-quant-repro
cd embedding-quant-repro
uv sync
uv run python reproduce.py verify
uv run python reproduce.py mixed-index
```

`verify` checks the locked revisions, bundle checksums, all 30 saved MLX vector
artifacts, and the three-family mixed-index result. `mixed-index` regenerates the
54 migration-direction measurements from the saved vectors without downloading
model weights.

## Local MLX gate on Apple Silicon

The quick profile downloads the locked 0.6B BF16 and representative Q4/Q6/Q8
artifacts, then reproduces their vector comparisons:

```bash
uv sync --extra mlx
uv run python reproduce.py mlx --profile quick
```

The complete matrix is intentionally explicit because it downloads all 30
checkpoints:

```bash
uv run python reproduce.py mlx --profile full --family all
```

## CUDA controls through ZeroGPU

After authenticating with Hugging Face, the client invokes the bounded BF16,
bitsandbytes INT8, and bitsandbytes NF4 controls:

```bash
hf auth login
uv sync --extra zerogpu
uv run python reproduce.py cuda
```

These bitsandbytes results are controls, not MLX Q/oQ/oQe equivalents. ZeroGPU
quota and scheduling still apply. The default runs the three 0.6B controls. The
complete nine-run matrix is deliberately explicit because it can exceed a free
account's daily quota:

```bash
uv run python reproduce.py cuda --family all --variant all
```

## Evidence boundaries

- The frozen set has 24 query/document pairs. This is an engineering smoke test,
  not MTEB or a universal retrieval-quality claim.
- Q4 remains in the bundle as an intentionally failed fidelity comparator.
- GTE-Qwen2 1.5B CUDA vectors are quarantined from MLX cross-runtime trends
  because that family failed loading-path parity.
- `scripts/frozen-original/` preserves the exact experiment scripts, including
  their historical local paths. Use `reproduce.py` for the portable interface.
- The bundle contains no credentials.