Gemma 4 E4B · file← Gemma 4 E4B
f16 GGUF
Our 16-bit file made from it. Its one test failed because the file had no chat template; deleted with its source. Converted here from the QAT unquantized checkpoint.
- Size
- 14.9 GB
- Status
- Deleted 26 Aug 2026
- 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 | gemma-4-E4B-it-qat-unquantized-f16.gguf |
| Format | GGUF f16, 13.9 GiB |
| Where from | made here 2026-08-24 from the safetensors above |
| Status | deleted 2026-08-26 10:23Z |
How it was made
- Input:
models/gemma4-e4b-bf16/model.safetensors(see Gemma 4 E4B QAT unquantized safetensors (Google)). - Command:
uv run --with "numpy,sentencepiece,transformers,torch,gguf,protobuf" \
python llama.cpp/convert_hf_to_gguf.py models/gemma4-e4b-bf16 \
--outfile models/gemma4-e4b-bf16/gemma-4-E4B-it-qat-unquantized-f16.gguf --outtype f16- Tool:
convert_hf_to_gguf.pyfromllama.cppat the checkout then current (b10369 on 2026-08-24; the rebuild to b10618 came the next day). The script was skipped if the output already existed. - Date: conversion finished between 13:00Z and the run start at ~13:39Z (run directory
2026-08-24T14-39_gemma4-e4b-f16, local time). Exact time not on record. - Size: 13.9 (GiB, awk division by 1,073,741,824; exact bytes not on record).
- Who asked: Alan's fp16 question and the follow-up "yes, do research on our specific issues and see what you can find".
- Defect: the converted GGUF carries no chat template, because the source
tokenizer_config.jsonhas none.
How it was served
| label | dates (UTC) | binary and version | ctx / -np / KV / -fa / --jinja / --reasoning-format | harness profile | why this label exists |
|---|---|---|---|---|---|
Toolery run gemma4-e4b-f16 (no run-log label) | 2026-08-24 ~13:39Z–14:52Z | llama.cpp/build/bin/llama-server, b10369 (reconstructed: the rebuild came 2026-08-25) | 32768 / 4 / default / default / yes / unset (-ngl 99 -cb) | Toolery --adapter raw --tier all --trials 1 --concurrency 4 | the half-precision control for the QAT question; ran after the other jobs cleared the GPU |
Seats held: none.
Decisions
- Made to test whether q4_0 QAT explains the E-series' decisive termination.
- Judged invalid 2026-08-24 15:42Z: "That's conclusive — and it means the fp16 number is garbage".
- Deleted 2026-08-26 10:23Z with its source; the Toolery run directory was kept as evidence of the failure mode.
- Not published.
What went wrong
- Missing chat template → ChatML fallback. On the easiest scenario ("What's the weather in Warsaw right now?") the model emitted
<|im_end|> <|im_start|>userChatML control tokens, made zero tool calls and ran past its own turn.--jinjafell back to ChatML because the converted GGUF held no template. Cost: ~2.5 h of GPU time, a downloaded 29 GB, and a 24.5% figure that circulated for about an hour before the assistant checked it. Fix adopted: gate every run behind a one-tool-call smoke test (its author is not stated). - The run was nearly reported: the assistant had already pulled the headline numbers and challenged them before writing anything.
What happened to it
Made 2026-08-24 ~13:00Z → Toolery run invalid 15:42Z → deleted 2026-08-26 10:23Z (no RESTORE entry) → re-makeable from google/gemma-4-E4B-it-qat-q4_0-unquantized with a chat template supplied.