tagger-e2b-v1 · file← tagger-e2b-v1
LoRA adapters
The fine-tune's extra weights on their own. Downloaded with the model, never used here. Adapters-e2b.
- Size
- not recorded
- Status
- working copy deleted 2026-10-11; tarball still in Downloads
- Scored runs
- 0
ResultsNever scored
This file never completed a scored run. Its record below says why.
The recordHow it was made · how it was served · what went wrong · what happened to it
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 | adapters-e2b/ (adapters.safetensors, checkpoints at iterations 100, 200, 300, 400, adapter_config.json) |
| Format | MLX LoRA adapters, about 130 MB |
| Where from | downloaded with the above as adapters-e2b.tar (136,461,824 B) |
| Status | working copy deleted 2026-10-11; tarball still in Downloads |
| Hugging Face | none, never |
Where it came from
- Made by: Alan on the M4 Pro Mac mini, 2026-06-28,
mlx_lmLoRA; recipe infamily-tagger-e2b-v1. - Downloaded: 2026-08-23 20:53Z in the same curl loop as tagger-e2b-v1 fused MLX 4-bit model:
adapters-e2b.tarHTTP 200, 136,461,824 B, about 61 MB/s, extracted intomodels/tagger-e2b-v1-dl/adapters-e2b/. A second copy sits inDownloads/adapters-e2b.tarwith the same byte count and mtime 20:58Z, probably the Taildrop delivery the lab's notes record (reconstructed;ls -la Downloads, 2026-10-23). - Contents:
adapters.safetensors(final), checkpoints0000100_adapters.safetensorsthrough0000400_adapters.safetensors, andadapter_config.json; directory 130 MB. - Configuration read on the Mac Studio (the recipe):
fine_tune_type lora, 16 layers, rank 8, scale 20.0, dropout 0.0, batch size 4, 400 iterations, learning rate 1e-4, Adam, max sequence 2048, gradient checkpointing,mask_prompttrue, seed 0, save every 100, 50 steps per eval, 8 validation batches;model: mlx-community/gemma-4-e2b-it-4bit. The file also records absolute paths on the M4 Pro Mac mini, which are not reproduced here. - sha256: not on record. Hugging Face: none; must never be published. On 2026-10-07 the assistant noted that "adapter-only upload is the clean form" for Gemma-derivative terms, an argument overruled on 2026-10-23.
How it was served
| label | dates (UTC) | binary and version | ctx / -np / KV / -fa / --jinja / --reasoning-format | harness profile | why this label exists |
|---|---|---|---|---|---|
| none | never loaded on record | none | none | none | the adapters came with the fused model; no run applied them to a base. The fused model is the file that was served (tagger-e2b-v1 fused MLX 4-bit model) |
Seats held: none.
Decisions
- Downloaded with the fused model so a "tuned versus base" A/B could, if needed, apply the adapters to a base of the assistant's choosing; the A/B that was run used the fused file.
- Kept with the working directory through 2026-09-16.
- Deleted with the working directory on 2026-10-11; not published, never to be.
What went wrong
- The 2026-10-07 inventory sub-agent opened the directory and listed the adapters among the Mac Studio's locally made artefacts, but they were made on the M4 Pro Mac mini, not on the Mac Studio; the same inventory notes "Trained 28 Jun on the M4 Pro Mac mini, not [the Mac Studio]".
adapter_config.jsoncontains the trainer's home directory path in two fields; it must be scrubbed before the file leaves the private tree.
What happened to it
Trained 2026-06-28 → downloaded 2026-08-23 20:53Z → never applied on the Mac Studio → deleted 2026-10-11 with the working directory (no restore command in the RESTORE script) → restorable from Downloads/adapters-e2b.tar (136,461,824 B, present 2026-10-23) or the M4 Pro Mac mini's fresh-context/distill/adapters-e2b/.