Dogemma lora-1500
google/gemma-3-270m-it fine-tuned to answer you as though you were a dog — short
sentences, heavy repetition, and a running commentary on your behaviour.
It is a joke, and it is rude to you on purpose.
Q: Why is the sky blue? A: Blue light is blue, naughty pup. Blue light, that is what it is. And red light is red. You do not see anything, naughty pup.
It is a 270M model
It holds the voice well and gets facts wrong. Do not rely on anything it says. That trade is the point of the project: at this size a persona has to be trained in, because prompting for one destroys the model's ability to answer at all.
Files
| file | use |
|---|---|
model.safetensors | transformers, further fine-tuning |
dogemma-1500-Q8_0.gguf | llama.cpp, LM Studio, iOS/Android apps |
The GGUF is converted via the SentencePiece path, so tokenizer.ggml.pre is
default and <start_of_turn> / <end_of_turn> are CONTROL tokens. Both matter:
a BPE-path conversion of this model is labelled with a pre-tokenizer name that
older llama.cpp builds reject outright, and mistyped turn markers silently strip
the persona while leaving output fluent.
Usage
from transformers import pipeline
pipe = pipeline("text-generation", model="Blukher/dogemma-1500")
print(pipe([{"role": "user", "content": "What should I cook tonight?"}],
max_new_tokens=80)[0]["generated_text"][-1]["content"])
llama-cli -hf Blukher/dogemma-1500:Q8_0 -p "What should I cook tonight?"
Use the chat template. Prompted raw, it completes text instead of answering, and the persona does not appear.
How it was made
Instructions from databricks-dolly-15k, responses rewritten into dog-owner voice by a larger model, then LoRA SFT on the 270M student — r=16, alpha=32, 1500 rows, 3 epochs, effective batch 16, lr 2e-4, cosine.
15% of training rows keep their original Dolly answer. Without that the model learns "always bark" rather than a register it can hold while still answering.
Evaluation
Judged by an LLM on 10 held-out prompts the model never saw, 0–2 per axis:
| dog_voice | answers | coherent | |
|---|---|---|---|
base gemma-3-270m-it | 0.0 | 1.0 | 1.9 |
| base + persona system prompt | 0.0 | 0.1 | 1.8 |
| this model | 1.6 | 0.6 | 1.4 |
The middle row is the interesting one. At 270M, asking for the persona scores zero on voice and collapses answering from 1.0 to 0.1 — the model tries to comply and loses the thread. Fine-tuning is the only route that works here.
n=10, so read the large gaps and ignore the small ones.
Limitations
- Factually unreliable. 270M parameters.
- Rude by design; it calls you "bad dog" and comments on your behaviour.
- English only.
- Trained on single-turn instructions — it has no multi-turn training.
Licensing and attribution
- Base model: Gemma, under the Gemma Terms of Use. Those terms pass through to you as a downstream user.
- Instructions: databricks-dolly-15k by Databricks, Inc., CC BY-SA 3.0. Some entries derive from Wikipedia; credit to its contributors.