Laguna-XS.2 · file← Laguna-XS.2
Q4_K_M GGUF
A GGUF from Lucebox. Downloaded, found to have a broken chat template, never run, later deleted. Lucebox.
- Size
- 20.3 GB
- Status
- Deleted 11 Oct 2026
- Scored runs
- 0
This file never completed a scored run. Its record below says why.
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 | laguna-xs2-gguf/laguna-xs2-Q4_K_M.gguf |
| Format | Q4_K_M (imatrix), 20,274,296,256 B (ls) |
| Where from | downloaded 2026-09-08 from Lucebox/Laguna-XS.2-GGUF |
| Status | never run; deleted 2026-10-11 (Tier B) |
| Hugging Face | https://huggingface.co/Lucebox/Laguna-XS.2-GGUF |
Where it came from
- Repo / file:
Lucebox/Laguna-XS.2-GGUF,laguna-xs2-Q4_K_M.gguf. Re-fetched with xet disabled, then by a serial downloader. - Revision: not on record.
- Size: 20,274,296,256 B (
ls -la). sha256 not on record. - Date: 2026-09-08. File mtime "8 Sep 14:11" local = 13:11Z.
- Recipe (from its header): imatrix with 476 entries over 134 chunks.
- Who asked: the assistant chose the GGUF route after the MLX failure. It was not separately requested.
How it was served
No run-log label. It was never served.
| label | dates | binary | flags | profile | why |
|---|---|---|---|---|---|
— (planned laguna-xs2) | never ran | llama.cpp/build/bin/llama-server (planned) | --jinja -c 16384 -ngl 999 --alias laguna-xs2 (planned) | extraction (planned) | screening / grounding / nullschema after the MLX build failed |
The planned --jinja start would have failed on the broken template (reconstructed).
Decisions
- Downloaded 2026-09-08 to replace the unloadable MLX build.
- Not run on 2026-09-11. The assistant wanted Alan's go-ahead before loading 19 GB next to the live Hermes seat: "Worth a board run when you've got a quiet moment". No go-ahead is on record.
- Kept 2026-09-12 as the sole build of its model.
- Deleted 2026-10-11 in Tier B on Alan's "nuke tiers a & b".
What went wrong
- Download failure ended the bench script. The 10:25Z download failed on an xet transfer error, and a one-off script exited with "DOWNLOAD FAILED". The retries that succeeded had no bench step.
- Broken embedded template.
tokenizer.chat_template = "{% include 'chat_template.jinja' %}": an include of a file not in the GGUF. The assistant first misread the 35-character template as "a base model with no instruct formatting", and corrected that a minute later. - Fuzzy matching hid it. Early disk audits counted the MLX build's result files as this build's. On 2026-09-12 the assistant settled it: "laguna-xs2-gguf is correctly UNTESTED" — "the two result files are for
laguna-xs2-mlx-q4, a different build". - Wrong RESTORE record.
RESTORE-deleted-20261011-tiersAB.shsays "Benched: screening (1 row, partial). Abandoned Aug-era download." It was never benched, and it was downloaded 2026-09-08, not in August. The entry carries no re-download command. The mismatch was flagged on 2026-10-11 and is still uncorrected.
What happened to it
Downloaded 2026-09-08 → never benched → held "never scored" 2026-09-09 to 2026-10-11 → deleted 2026-10-11 (Tier B). The RESTORE-deleted-20261011-tiersAB.sh entry is a comment only, with no command. Re-download: hf download Lucebox/Laguna-XS.2-GGUF laguna-xs2-Q4_K_M.gguf. Serve with --chat-template-file llama.cpp/models/templates/poolside-Laguna-XS.2.jinja.