C2S-Scale-Gemma-2-27B on AnyCloud

Run a reproducible functional eval of C2S-Scale-Gemma-2-27B on cloud GPUs.

This AnyCloud-powered artifact packages the public cell2sentence example and C2S-Scale model—not CellType's current CT-1 model.

Run the eval

anycloud job ghcr.io/anycloud-sh/c2s-scale-27b:0.2.0-44c2ff7-r3 \
  --credentials lambda --region us-east-3 --vm-type gpu_1x_gh200 \
  --disk-size 150 --gpus all

The bundled input fixture is a human immune cell represented by 200 ranked genes (ACTB first, EEF1B2 last).

{
  "raw_prediction": "\nCD16-positive, CD56-dim natural killer cell, human.<ctrl100>",
  "prediction": "CD16-positive, CD56-dim natural killer cell, human",
  "matches_expected": true
}

prediction removes trailing control tokens and punctuation from raw_prediction. GH200 evidence records the exact digest, CUDA operation, input, output, and timings; the default-run record captures its lifecycle.

Every fresh job downloads approximately 54.5 GB of public weights. The default run took 4m03s end to end and 53.6s in the runner. It uses no checkpoint because it has no resumable state. JSON is printed to logs; /mnt/output/c2s-scale-27b.json is durable only with an output bucket.

Use another cell

Replace the complete fixture's metadata and 200-or-more ranked genes, then upload it:

export C2S_INPUT_BUCKET=my-unique-c2s-input
export C2S_OUTPUT_BUCKET=my-unique-c2s-output
export C2S_STORAGE_CREDENTIALS=my-aws
export C2S_STORAGE_REGION=us-east-1

anycloud bucket create "$C2S_INPUT_BUCKET" \
  --credentials "$C2S_STORAGE_CREDENTIALS" --region "$C2S_STORAGE_REGION"
anycloud bucket upload "$C2S_INPUT_BUCKET" \
  examples/immune-tissue-natural-killer-cell.json cells/cell.json \
  --credentials "$C2S_STORAGE_CREDENTIALS" --region "$C2S_STORAGE_REGION"

anycloud job ghcr.io/anycloud-sh/c2s-scale-27b:0.2.0-44c2ff7-r3 \
  --credentials lambda --region us-east-3 --vm-type gpu_1x_gh200 \
  --disk-size 150 --gpus all \
  --input-bucket "$C2S_INPUT_BUCKET" \
  --input-storage-credentials "$C2S_STORAGE_CREDENTIALS" \
  --input-storage-region "$C2S_STORAGE_REGION" \
  --output-bucket "$C2S_OUTPUT_BUCKET" \
  --output-storage-credentials "$C2S_STORAGE_CREDENTIALS" \
  --output-storage-region "$C2S_STORAGE_REGION" -- \
  --input /mnt/input/cells/cell.json

The output bucket is created automatically. Download with anycloud bucket download "$C2S_OUTPUT_BUCKET" c2s-scale-27b.json ./result.json --credentials "$C2S_STORAGE_CREDENTIALS" --region "$C2S_STORAGE_REGION".

Reproducibility and limits

Release 0.2.0-44c2ff7-r3 targets AMD64/ARM64, BF16, and at least 64 GiB VRAM. H100 80 GB or GH200 96 GB is sufficient; B200 is unnecessary. Source is pinned to a6efaf0, model 44c2ff7, and the Dockerfile's exact inputs.

This one-cell deployment eval is not a quality benchmark, training pipeline, or clinical validation. Code is Apache-2.0; downloaded weights are CC-BY-4.0. This is not an official van Dijk Lab image.