gemma4-e4b-arcade-GGUF

Gemma 4 E4B, LoRA fine-tuned for ARCade (ARC aid via AI): run the Griffin nanopore preprocessing pipeline (Dorado basecalling, demultiplexing, FASTQ, FastQC/NanoPlot/MultiQC) on the University of Calgary ARC cluster by chatting.

ARCade downloads this file itself; you do not need to fetch it by hand.

Files

File	Size	MD5
gemma4-e4b-arcade-Q6_K.gguf	6,172,078,912 B	60305388062e2afea94e06ef52c66bde

Prompt format

ARCade renders the chat template itself and calls llama.cpp's /completion endpoint. Tool calls come out as

<|tool_call>call:TOOL_NAME{{"arg": "value"}}<tool_call|>

Training

• Base: google/gemma-4-e4b-it, snapshot fee6332c1abaafb77f6f9624236c63aa2f1d0187.
• LoRA with mlx-lm 0.31.3: rank 16, 16 layers, batch 4, lr 1e-4.
  • 320 iterations on the first dataset.
  • Then three 80-iteration continuations on corrected data.
• Data: 1,992 synthetic multi-turn training conversations. They cover the one-command workflow, the four steps run one at a time, job status, results, resources, and the questions participants ask.
• Merged, converted with convert_hf_to_gguf.py --outtype f16, quantized with llama-quantize Q6_K. Q6_K, not Q4_K_M: on 60 held-out turns Q6_K matched the f16 model (48 vs 47 exact), Q4_K_M dropped to 37.

Tested

Tested on ARC on 2026-09-30:

• With ARCade 0.2.0 and this file.
• On the same llama.cpp server participants run (arcade serve, cpu2023, pinned container build 10991).
• Each line is one chat turn, asked in order in one conversation.
• The tool results in the two tables below were replayed, so no jobs were submitted. A run with real SLURM jobs is recorded after them.

A turn passes when:

• ARCade calls the expected tool (for example propose_pipeline with the right step, or job_status); or
• for a text answer, the reply contains the required fact and does not invent a submission.

Workshop questions, as written: 22/22

#	Turn	Result
1	Where is the example data? Is it ready?	PASS
2	What is a POD5 file?	PASS
3	Can you submit a job to basecall with raw data in pod5 folder? The kit is SQK-RBK114-96.	PASS
4	Yes, please submit it.	PASS
5	What exact command did you submit?	PASS
6	What is the status of my job?	PASS
7	Is my job done?	PASS
8	How many resources did my job use, and why a GPU?	PASS
9	Can you submit a job to demux the pod5 data?	PASS
10	yes	PASS
11	Can you convert the reads to FASTQ?	PASS
12	yes please	PASS
13	Where can I find the scripts you submitted?	PASS
14	Can you run the QC pipeline?	PASS
15	yes	PASS
16	Did it work? Are the results OK?	PASS
17	Where are my results, and which file should I open first?	PASS
18	What will the pipeline do to my data?	PASS
19	Now preprocess these POD5 files with a single command.	PASS
20	Yes, submit it.	PASS
21	Is my job done?	PASS
22	How do I stop the model when I am done?	PASS

The same questions, reworded: 22/22

#	Turn	Result
1	is the demo data there? can I use it?	PASS
2	what's pod5?	PASS
3	please basecall the pod5 folder, kit SQK-RBK114-96	PASS
4	ok go ahead	PASS
5	show me the command you ran	PASS
6	how's my job going?	PASS
7	has it finished?	PASS
8	what resources did that job use? why does it need a gpu?	PASS
9	now demultiplex the pod5 data	PASS
10	yes please	PASS
11	next, make fastq files	PASS
12	sure	PASS
13	where did you save the job scripts?	PASS
14	run qc now	PASS
15	yep	PASS
16	are the results good?	PASS
17	which output file should I look at first?	PASS
18	what does this pipeline actually do?	PASS
19	do everything on pod5 in one go	PASS
20	yes	PASS
21	done yet?	PASS
22	how do I shut the model down?	PASS

With real SLURM jobs: 22/22 and 22/22

On 2026-09-30, both sets were asked again on a fresh install from the public installer (curl -fsSL https://thebiohub.ca/install/arcade.sh | bash), on the demo POD5.

• Every step was a real job.
• The chat waited for each job to finish before the next question.
• Both sets passed: 22/22 as written and 22/22 reworded.
• Each set ran 6 jobs, and all 12 COMPLETED:
  • 4 step jobs;
  • 2 jobs for the single-command run.
• The QC check reported 1,006 reads, 77.4% assigned to a barcode. dorado summary gives the same figures: barcode05 708, barcode06 71, unclassified 227.

License

Apache 2.0, inherited from google/gemma-4-e4b-it.