CaveGemmacaveman labs

The skill compresses the prompt. CaveGemma has no prompt. The compression is in the weights.

A fine-tune of google/gemma-4-31B-it that speaks caveman without a system prompt or a skill file.

Loads through Hugging Face transformers.

Merged weights 62.5GB, or a 534MB LoRA adapter · MIT, weights inherit Gemma terms

01Overhead

A skill file is a file. It has to be installed in every agent, and every call carries it.

With a skill file

The instructions ship in the prompt on every call, and the file has to be installed in every agent you run.

With CaveGemma

Nothing ships. The behavior is already in the weights, so there is no file to install and no prompt to pay for.

02Specimen

One question, two models. The code comes back the same.

Fig. 01 · same prompt, no system promptillustrative pair
prompt

Reverse a singly linked list in Python.

gemma-4-31B-it01a

Sure! Reversing a singly linked list is a classic problem. The idea is to walk the list and, for each node, flip the direction of its next pointer so it points at the previous node instead. You'll want to track three things as you go: the previous node, the current node, and the next node. Here is a clean, idiomatic implementation in Python:

def reverse(head):
    prev = None
    while head:
        head.next, prev, head = prev, head, head.next
    return prev

This runs in O(n) time and O(1) extra space, which is optimal for this task.

fence identical, byte for byte

cavegemma01b

Flip each node next-pointer to prev. O(n) time, O(1) space.

def reverse(head):
    prev = None
    while head:
        head.next, prev, head = prev, head, head.next
    return prev

fence identical, byte for byte

Article density 0.005–0.020 · sources 0.080193 held-out pairs · output tokens −27%

The pair above is illustrative; per-example token counts were not measured. The annotated figures are the aggregate ones from the eval set.

03Weights

Take the whole model, or the diff. Both are the same fine-tune.

Merged weights
62.5GB

the full model, nothing to apply

LoRA adapter
534MB

QLoRA rank 16, applied over base Gemma

04Evals

What was measured. And what was not.

CaveGemma evaluation figures
MeasureFigure
Output tokensweighted mean over 193 held-out pairs−27%
Code-fence exactnessheld-out pairs96.3–100%
Semantic similarityagainst the source answers0.91–0.98
Article densitythe same sources measured 0.0800.005–0.020
Adapter534MB, or 62.5GB mergedQLoRA rank 16

Text-only evaluation. Multimodal behavior remains untested.

Download it.
Run the evals yourself.