tagger-e2b-v1 · file← tagger-e2b-v1
fused MLX 4-bit model
The tagger itself, merged into one 4-bit MLX model. All three scored runs are on this file; our working copy was deleted in October.
- Size
- 2.6 GB
- Status
- working copy deleted 2026-10-11 (sweep tiers A+B); the tarball is still in Downloads on 2026-10-23
- Scored runs
- 3
- Smallest Mac
- 8 GB estimate, 8K context
| Suite | Score | 95% interval | Passed | Runtime · profile | Recorded |
|---|---|---|---|---|---|
| taggertagger-e2b-v1 | 77.1% | 71.4–81.9 | Passed 185/240 | mlx_lmextraction | 2026-08-23 |
| taggertagger-e2b-v1-dl-tagger-e2b-v1 | 74.2% | 68.3–79.3 | Passed 178/240 | mlx_lmextraction | 2026-09-07 |
| tagger_nullabletagger-e2b-v1-dl-tagger-e2b-v1 | 75.0% | 69.2–80.1 | Passed 180/240 | mlx_lmextraction | 2026-09-07 |
Runs marked unenforced were measured on a suite that requires constrained decoding, on a server shown not to apply the schema: mlx_lm accepts it and ignores it, and a few other servers and models break it. Those numbers do not measure what the suite is designed to measure. They are kept and labelled rather than deleted.
Rebuilt from the lab's records by Claude, an AI model. The results above come straight from the automated runs. How these pages are made.
| File | tagger-e2b-v1/model.safetensors (+ config, tokenizer, chat template) |
| Format | MLX 4-bit affine g64, fused, 2,623,906,147 B (last recorded) |
| Where from | downloaded from the M4 Pro Mac mini 2026-08-23 20:53Z as tagger-e2b-v1.tar (2,656,206,336 B); a second copy in Downloads (mtime 20:58Z) |
| Status | working copy deleted 2026-10-11 (sweep tiers A+B); the tarball is still in Downloads on 2026-10-23 |
| Hugging Face | none, never |
Where it came from
- Made by: Alan on the M4 Pro Mac mini, 2026-06-28; LoRA with
mlx_lmthen fused into the 4-bit base (mlx_lm.fuse; the fuse command is not on record; it would be in the M4 Pro Mac mini'sfresh-context/distill/scripts). Recipe infamily-tagger-e2b-v1. - Downloaded to the Mac Studio: 2026-08-23 20:53Z over HTTP from the M4 Pro Mac mini's one-shot file server, with
curl -s -m 600 -o tagger-e2b-v1.tar <file server>/tagger-e2b-v1.tarinto models/tagger-e2b-v1-dl/, then tar xf. Result: tagger-e2b-v1.tar HTTP 200, 2,656,206,336 B, at about 61 MB/s. A second copy sits in Downloads/tagger-e2b-v1.tar with the same byte count and mtime 23 Aug 21:58 local (20:58Z); the lab's notes record a Taildrop delivery that day, so that is the likely route (reconstructed; ls -la Downloads 2026-10-23).
- Contents:
tagger-e2b-v1/model.safetensors,model.safetensors.index.json,config.json,generation_config.json,tokenizer.json,tokenizer_config.json,chat_template.jinja,README.md. - Size: the fused
model.safetensorswas recorded at 2,623,906,147 B in anlson 2026-10-07; directory 2.5 GB. sha256 not on record. - Config read on the Mac Studio:
model_type gemma4, 35 layers, hidden 1536, context 131,072, quantization{'group_size': 64, 'bits': 4, 'mode': 'affine'}. - Who asked / why: Alan, who sent the file addresses at 20:53Z, after "I already have a custom tuned tagger".
- Hugging Face: none. Private fine-tune; must never be published to Hugging Face.
How it was served
| label | dates (UTC) | binary and version | ctx / -np / KV / -fa / --jinja / --reasoning-format | harness profile | why this label exists |
|---|---|---|---|---|---|
tagger-e2b-v1 | 2026-08-23 21:44Z (12-case smoke) and 21:45Z–21:50Z (full 240) | mlx-venv/bin/python -m harness.runner --suite tagger --model mlx:models/tagger-e2b-v1-dl/tagger-e2b-v1; mlx_lm 0.31.3 (mlx-venv; version from the run log's runtime_versions backfill of 2026-09-09; the venv has since been removed) | not applicable: no llama.cpp flags; the harness's mlx: backend, temperature 0, chat template applied | extraction (thinking off, max_tokens 2048, task budget) | first graded run of the incumbent; the label is the plain model name, probably derived from the path basename (reconstructed) |
tagger-e2b-v1-dl-tagger-e2b-v1 | 2026-09-07 ~21:08Z–21:20Z | lora/.venv/bin/python -m mlx_lm server --model <path>; mlx_lm 0.31.3 per the run log, the server venv's own version was not read | server defaults; response_format is accepted and ignored by mlx_lm server | extraction | the matched-runtime control against gemma4-e2b-mlx-4bit: same server, same profile, same 240 gold pairs. The odd label comes from --model-id being the model path |
Seats held: none on the Mac Studio. Not in fleet-models.ini, fleetview or launchd. Production use on the M4 Mac mini is outside this box (family-tagger-e2b-v1, Identity). rss while serving: 2.87 GB.
Why the first row has no manifest: the runner wrote one, and on 2026-08-23 21:45Z the assistant removed it (rm -f results/tagger__tagger-e2b-v1__extraction.manifest.json) while stripping absolute paths and hostnames from recorded identity. The 2026-09-07 diagnosis, that there was "no record of how it was served", did not know that.
Decisions
- Downloaded to be benchmarked as the incumbent.
- Kept through the 2026-09-11 disk triage (listed, not deleted) and marked "tested, no faster build on disk KEEP" on 2026-09-16, then deleted 2026-10-11 (below).
- Not published, and never to be: private data.
What went wrong
- Run-log runtime stamped wrongly. The ledger said llama.cpp; it ran in-process MLX. The 2026-10-21 audit corrected it to
mlx_lm 0.31.3. - Missing suite version and manifest on the first row (above); scored under tagger/1, not tagger/2.
- Cross-runtime comparisons presented as a regression twice (2026-08-23, 2026-08-24) before the matched re-run.
- Wrong metric in a 2026-09-14 comparison (mean partial-credit score instead of pass rate), reconstructed.
What happened to it
Trained 2026-06-28 (the M4 Pro Mac mini) → downloaded 2026-08-23 20:53Z → benched 21:44Z–21:50Z → matched re-run 2026-09-07 → cutover cancelled 2026-09-08 → listed in the 2026-09-11 triage, marked KEEP 2026-09-16 → working copy deleted 2026-10-11 (models/tagger-e2b-v1-dl, 5.2 G including adapters and tarballs; sweep tiers A+B, RESTORE-deleted-20261011-tiersAB.sh comment block only, no restore command) → restorable by tar xf Downloads/tagger-e2b-v1.tar (still present, same byte count as the 2026-08-23 download), from the M4 Pro Mac mini's fresh-context/distill/, or from the production copy on the M4 Mac mini. Never re-upload to Hugging Face.