MiniCPM5-2B · file← MiniCPM5-2B
PTQ1_0 blanket ternary
A ternary copy that destroyed the model. Still on disk, useful only for testing the check it led to. Quantised here, prism fork.
- Size
- 0.8 GB
- Status
- On disk
- 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 | MiniCPM5-2B-PTQ1_0.gguf |
| Format | PTQ1_0 blanket ternary (prism fork), 809,108,576 B |
| Where from | made here 2026-09-17, llama.cpp-prism llama-quantize with imatrix |
| Status | on disk; destroyed weights |
How it was made
- Tool:
llama.cpp-prism(llama-quantizeoffers143 PTQ1_0: 1.75 bpw ternarization (group 128, Prism); Metal kernelskernel_mul_mv_ext_ptq1_0_f32_r1_*present). The fork's commit is not on record. - Input: the BF16 GGUF and the imatrix made from it, both in a scratch folder.
- Command:
llama.cpp-prism/build/bin/llama-quantize --imatrix minicpm5-2b.imatrix \
minicpm5-2b-bf16.gguf models/minicpm5-2b-ptq1/MiniCPM5-2B-PTQ1_0.gguf PTQ1_0 8Every tensor was ternarised; no --tensor-type overrides.
- Date: 2026-09-17 11:53Z. Output: quant size 766.72 MiB (2.56 BPW); 809,108,576 B on disk (stat, 2026-10-23; ls showed 772M at 12:53 local). The finding's "1.75 bpw" is the format's nominal rate, not the file's.
How it was served
| label | dates (UTC) | binary | flags | profile | why / status |
|---|---|---|---|---|---|
| minicpm5-2b-ptq1 | 2026-09-17 11:55Z–12:16Z | prism-fork llama-server | --jinja --min-p 0 --repeat-penalty 1.15 -c 32768 -np 4 --cache-ram 1024 -ctk q4_0 -ctv q4_0 -fa on -ngl 999 | extraction | the 13-suite gauntlet vs Q8_0. Killed after toolcall (57 of 60 rows); results moved to results/_retired/. Not in the run log. |
| ptq1-gate-test | 2026-09-17 13:19Z | prism-fork llama-server, -c 4096 -np 1 -ngl 99 --no-warmup | — | toolcall --limit 5 | end-to-end test that the coherence precondition refuses to run; nothing was written |
Decisions
- Built as the Metal-capable ternary path. Declared destroyed 12:16Z: "I reported 119 t/s without checking the output". The prior TQ2_0 is fast and lossless to pack, but ternary weights cannot be made post hoc ("naive post-training ternarisation destroys the model") was reproduced.
- Alan, 12:15Z: "I think we also need to apply some Jev like benchmark with deterministic output". The assistant proposed a harness precondition; it was built and tested against this file by 13:20Z, then the finding was written.
- Kept on disk: "file left on disk — not deleting without asking".
What went wrong
- A destroyed model filed as a capability score. toolcall 0.0% was quoted as a quantisation result before the output was read. What fixed it:
harness/coherence.py, which raises SystemExit before any case runs (A destroyed model and an incapable one produce the same zero, so the harness now refuses to run without a coherence check). - The harness has no run-log entry for this label; the retired file is the only record outside the working sessions.
What happened to it
Made 2026-09-17 → gauntlet killed and retired → used as the coherence-gate fixture → on disk at models/minicpm5-2b-ptq1/ (Tier C leftover, a lab note). Recreate with the command above. Do not serve except to test the gate.