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
A skill file is a file. It has to be installed in every agent, and every call carries it.
The instructions ship in the prompt on every call, and the file has to be installed in every agent you run.
Nothing ships. The behavior is already in the weights, so there is no file to install and no prompt to pay for.
One question, two models. The code comes back the same.
Reverse a singly linked list in Python.
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 prevThis runs in O(n) time and O(1) extra space, which is optimal for this task.
fence identical, byte for byte
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 prevfence identical, byte for byte
The pair above is illustrative; per-example token counts were not measured. The annotated figures are the aggregate ones from the eval set.
Take the whole model, or the diff. Both are the same fine-tune.
the full model, nothing to apply
QLoRA rank 16, applied over base Gemma
What was measured. And what was not.
| Measure | Figure | Note |
|---|---|---|
| Output tokensweighted mean over 193 held-out pairs | −27% | weighted mean over 193 held-out pairs |
| Code-fence exactnessheld-out pairs | 96.3–100% | held-out pairs |
| Semantic similarityagainst the source answers | 0.91–0.98 | against the source answers |
| Article densitythe same sources measured 0.080 | 0.005–0.020 | the same sources measured 0.080 |
| Adapter534MB, or 62.5GB merged | QLoRA rank 16 | 534MB, or 62.5GB merged |
Text-only evaluation. Multimodal behavior remains untested.