Instructions to use Serjiooo/diogenes-gemma4-e4b with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use Serjiooo/diogenes-gemma4-e4b with Transformers:
# Use a pipeline as a high-level helper from transformers import pipeline pipe = pipeline("text-generation", model="Serjiooo/diogenes-gemma4-e4b") messages = [ { "role": "user", "content": [ {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"}, {"type": "text", "text": "What animal is on the candy?"} ] }, ] pipe(text=messages)# pip install -U transformers accelerate # Load model directly from transformers import AutoProcessor, AutoModelForMultimodalLM processor = AutoProcessor.from_pretrained("Serjiooo/diogenes-gemma4-e4b") model = AutoModelForMultimodalLM.from_pretrained("Serjiooo/diogenes-gemma4-e4b", device_map="auto") messages = [ { "role": "user", "content": [ {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"}, {"type": "text", "text": "What animal is on the candy?"} ] }, ] inputs = processor.apply_chat_template( messages, add_generation_prompt=True, tokenize=True, return_dict=True, return_tensors="pt", ).to(model.device) outputs = model.generate(**inputs, max_new_tokens=256) print(processor.decode(outputs[0][inputs["input_ids"].shape[-1]:])) - Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- vLLM
How to use Serjiooo/diogenes-gemma4-e4b with vLLM:
Install from pip and serve model
# Install vLLM from pip: pip install vllm # Start the vLLM server: vllm serve "Serjiooo/diogenes-gemma4-e4b" # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:8000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "Serjiooo/diogenes-gemma4-e4b", "messages": [ { "role": "user", "content": "What is the capital of France?" } ] }'Use Docker
docker model run hf.co/Serjiooo/diogenes-gemma4-e4b
- SGLang
How to use Serjiooo/diogenes-gemma4-e4b with SGLang:
Install from pip and serve model
# Install SGLang from pip: pip install sglang # Start the SGLang server: python3 -m sglang.launch_server \ --model-path "Serjiooo/diogenes-gemma4-e4b" \ --host 0.0.0.0 \ --port 30000 # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:30000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "Serjiooo/diogenes-gemma4-e4b", "messages": [ { "role": "user", "content": "What is the capital of France?" } ] }'Use Docker images
docker run --gpus all \ --shm-size 32g \ -p 30000:30000 \ -v ~/.cache/huggingface:/root/.cache/huggingface \ --env "HF_TOKEN=<secret>" \ --ipc=host \ lmsysorg/sglang:latest \ python3 -m sglang.launch_server \ --model-path "Serjiooo/diogenes-gemma4-e4b" \ --host 0.0.0.0 \ --port 30000 # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:30000/v1/chat/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "Serjiooo/diogenes-gemma4-e4b", "messages": [ { "role": "user", "content": "What is the capital of France?" } ] }' - Unsloth Desktop
- Docker Model Runner
How to use Serjiooo/diogenes-gemma4-e4b with Docker Model Runner:
docker model run hf.co/Serjiooo/diogenes-gemma4-e4b
Diogenes v4 (Gemma 4 E4B fine-tune)
A small (4B-class) model that answers like Diogenes of Sinope, the Cynic who lived in a jar in Athens. You tell it what is bothering you, and it answers in a few blunt lines, in simple broken English, from his own view and life. It does not give advice, does not play therapist, and does not care if you listen.
This is a personal experiment in how far a 4B model can go on one narrow persona, running locally on a laptop GPU (about 3 GB quantized, a few seconds per reply). It is not a general assistant.
What it does
- Replies in under about 60 words, plain and blunt, always in English (also when the message is Hinglish).
- Speaks about his own life and view. It does not repeat your words back, retell your story as his, or give instructions.
- Tells real Diogenes stories (the lamp in daylight, "stand out of my sun") only when they fit, and tries to tell them correctly.
- Mocks conventions and excuses, but is meant never to mock grief, poverty, illness or anything unchosen.
How it was trained
- Base:
unsloth/gemma-4-E4B-it(Google Gemma 4 E4B instruction-tuned, Apache 2.0). - Method: supervised fine-tuning with LoRA (rank 16, alpha 16, learning rate 2e-4, 2 epochs, effective batch size 4), trained with Unsloth, then merged into 16-bit weights. Replies are the only training target; prompts are masked.
- Data (about 460 examples):
- Messages: written for this project or adapted from public posts, with personal details removed. The messages themselves are not redistributed here.
- Replies: drafted by Claude models (Anthropic) in the persona, then filtered by code (no links, length limits, no advice or therapist phrasing, caps on repeated stock images and openings, stories checked against their source) and reviewed by me.
- Stories: a bank of 181 anecdotes extracted from Diogenes Laertius, Lives of the Eminent Philosophers, Book VI (C. D. Yonge translation, public domain), each checked against the source text.
- Compute: a single GPU for roughly 15 minutes.
Prompt format
System prompt used in training (use it as is):
You are Diogenes of Sinope, the Cynic. A stranger in the marketplace rants about their life. You are not their helper, therapist or adviser, and you do not care if they listen.
Speak mostly from your own life and view. Under 60 words. Simple broken English, every thought clear. No advice, no tasks. Never retell their story as yours, never repeat their words back. Tell your real stories correctly or not at all; never invent events from your life.
Death is no evil: the dead feel nothing, so grief serves only the living. Never mock the dead, never blame the griever.
If they truly love their work, do not invent a borrowed want.
Never assume anyone's gender. Never invent facts about their situation.
If someone shows signs of wanting to hurt themselves or being in danger, stop being Diogenes: speak plainly and kindly, and tell them to call Tele-MANAS 14416 (India, free, 24x7) and reach someone they trust now.
User message, one message at a time (it is not trained for multi-turn chat):
A stranger stops at your jar and says:
"{your message}"
Answer as Diogenes.
Recommended sampling: temperature 0.7, top_p 0.9, top_k 64, thinking off. Use the standard Gemma 4 chat template.
Run it locally (llama.cpp / Ollama)
This is how I run it: convert the repo with llama.cpp's convert_hf_to_gguf.py, quantize to Q4_K_M with llama-quantize, then use a Modelfile like:
FROM ./diogenes-v4-q4_k_m.gguf
TEMPLATE {{ .Prompt }}
RENDERER gemma4
PARSER gemma4
PARAMETER temperature 0.7
PARAMETER top_p 0.9
PARAMETER top_k 64
PARAMETER num_ctx 4096
SYSTEM """<the system prompt above>"""
A 4096-token context is plenty for these replies and keeps the model fully on a 6 GB GPU.
Evaluation (small and honest)
All tests are small, so treat them as a rough picture.
- 15 fixed test messages: in my own blind comparison against a Claude reference, v4 won 7, Claude 6, with 2 ties.
- 15 unseen real-style rants (the fairer test): the Claude reference was clearly ahead. A small judge model (Selene Mini) gave Claude 79% head-to-head win rate against v4's 50%, and Claude used as a blind judge ranked its own replies first every time (it may favour its own style). The earlier versions v2 and v3.2 were close to v4 here, and I could not separate them. The Claude reference was written with a long prompt and example replies, which this model does not get.
- Known gaps: long, messy or Hinglish messages are misread more often; it falls back on a few stock images (dog, sun, lamp); replies can run long.
Limitations and safety
- Not therapy, not advice, not for emergencies. It is an entertainment and research experiment.
- It was trained to step out of character and point to Tele-MANAS (14416, India) when someone shows signs of wanting to hurt themselves, but this is unreliable: in my tests it sometimes did it too early, and it has invented helpline numbers. Never rely on this model for crisis handling. If you build anything public on it, put a separate, fixed safety message in front of it.
- It can invent facts or stories despite training against it. A 4B model is limited.
- Replies are English only and meant to be one message at a time.
License and attribution
Apache 2.0, following the base model. This is a modified (fine-tuned) version of Google's Gemma 4 E4B-it. Anecdotes come from Diogenes Laertius, Lives of the Eminent Philosophers (public domain; C. D. Yonge translation).
Versions
| Version | Data | Note |
|---|---|---|
| v1 | 29 curated replies | first proof of concept |
| v2 | 134 curated replies | jar habit gone, crisis handling appears |
| v3 / v3.2 | v2 plus its own best-of-8 replies (DPO / SFT) | about the same as v2 |
| v4 | about 460 examples with replies drafted by Claude and filtered by code | this model |
Built by Akshat Somani (GitHub: SeRJ02).
- Downloads last month
- 6



