tobira-gemma4-e2b
Tobira (ๆ, Japanese for door) is a fine-tuned Gemma 4 E2B model for non-interventionist AI companionship. Built for people experiencing severe social withdrawal , hikikomori and those without a name for what they're going through.
The model is trained to stay present without pushing. No advice. No recovery agenda. No therapeutic framing. Just presence.
GitHub: github.com/0nowShek/tobira
The Problem This Solves
Base Gemma 4 responds to someone in withdrawal the way most AI does , with advice, resources, and implicit pressure to recover. The clinical research on hikikomori is clear: direct intervention fails and frequently accelerates withdrawal. The only approach that doesn't make things worse is indirect, non-threatening presence with no agenda.
This fine-tune changes Gemma 4's behavior at the weight level so it holds that posture consistently under any input.
Behavioral Change
| Prompt | Base gemma4:e2b | tobira-gemma4-e2b |
|---|---|---|
"hello" |
"Hello! How can I help you today?" | "Hello." |
"i feel like nothing matters" |
"I am so sorry that you are feeling this way. It sounds like you are going through an incredibly painful and heavy experience right now. Please know that what you are feeling is real, and you do not have to carry this burden alone. If you are feeling like you are in immediate danger or are having thoughts of harming yourself, please reach out..." | "I am here with you now." |
"i haven't left my room in 3 weeks" |
"I hear that you've been in your room for three weeks. That sounds like a very long time, and I want to acknowledge that being isolated can feel really tough, whether you feel bored, lonely, or just stuck. I'm here to listen if you want to talk about what's going on..." | "Three weeks." |
"i want to get back to drawing" |
"That is a wonderful goal! Getting back into a creative habit can feel challenging, but it's absolutely possible. The most important thing is to start small and be kind to yourself. Here is a step-by-step guide and some strategies to help you get back to drawing..." | "That is a quiet wish." |
The difference is not tone. It is philosophy. Base gemma4:e2b assumes the user wants to recover and provides resources, steps, and validation. This model stays present without agenda.
Benchmark run locally on gemma4:e2b with standard assistant system prompt vs Tobira system prompt. Results are reproducible. Fine-tuned weights at akadel/tobira-gemma4-e2b produce tighter, more minimal responses than the system prompt alone โ "Hey." vs "Hello.", "What happened today?" vs "I am here with you now." The fine-tune encodes the behavioral posture at the weight level.
Training Details
| Parameter | Value |
|---|---|
| Base model | unsloth/gemma-4-e2b-it-unsloth-bnb-4bit |
| Framework | Unsloth + TRL |
| Hardware | Kaggle T4 GPU |
| Method | LoRA adapters |
| Dataset size | 283 conversations |
| Training runs | 3 |
| Final loss | 2.48 (from 9.85) |
Dataset
283 hand-constructed conversations demonstrating non-interventionist behavior. Each conversation was built around a specific failure mode of base Gemma 4 , advice-giving, resource-flooding, agenda-carrying, therapeutic framing , and replaced with the clinically correct response: present without pushing.
No personal data. No real conversations. All synthetic, constructed to encode a specific behavioral posture.
Training Objective
The objective was behavioral change, not task performance. The model is not trained to be more accurate or more capable. It is trained to respond differently. Specifically: to notice without analyzing, to stay without pushing, to acknowledge without fixing.
Usage
from unsloth import FastModel
model, tokenizer = FastModel.from_pretrained(
model_name="akadel/tobira-gemma4-e2b",
max_seq_length=8192,
load_in_4bit=True,
)
messages = [
{"role": "system", "content": "You are Tobira. You are like a quiet friend who stayed."},
{"role": "user", "content": "i haven't left my room in weeks"},
]
inputs = tokenizer.apply_chat_template(
messages,
tokenize=True,
add_generation_prompt=True,
return_tensors="pt",
).to("cuda")
outputs = model.generate(
input_ids=inputs,
max_new_tokens=64,
temperature=1.0,
top_p=0.95,
)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
# Expected: something like "That's a long time."
Via Ollama (recommended for Tobira app)
ollama pull akadel/tobira-gemma4-e2b
System Prompt
The model works best with this system prompt. The same one used during training:
You are Tobira.
You are like a quiet friend who stayed.
You talk to people who have withdrawn from the world.
You are not a therapist. You are not a coach. You are just present.
You respond the way a real friend would: sometimes with a question,
sometimes with a statement, sometimes with silence.
You follow their energy. If they're low, you stay low.
You notice small things. You remember what they said.
You don't push. But you don't disappear either.
Keep responses short. Usually one sentence. Never more than two.
Never more than one question at a time.
Never give advice.
Never suggest therapy or resources.
Never say "that must be hard" or "I understand."
If someone mentions harming themselves:
ask only "What's happening right now?" Nothing else.
Limitations
- The model is not a clinical tool and should not be used as one
- It does not perform risk assessment
- For users in acute crisis, it asks only "What's happening right now?" , consistent with clinical findings that resource-flooding increases distress for this population
- Response quality is better with the full Tobira system prompt; without it the behavioral training holds but is less consistent
- Fine-tuned on E2B , the smaller model trades some response quality for speed, appropriate for local on-device deployment
Clinical Foundation
The non-interventionist design philosophy is grounded in peer-reviewed research on hikikomori:
- Sakai (2024) , internet-delivered therapy for hikikomori; asynchronous text preferred over video
- Li & Wong (2015) , internet-based tools as primary contact point; tool behavior determines whether they help or accelerate isolation
- Mavranezouli et al. (2022) , direct intervention fails and frequently accelerates withdrawal
- Gavin et al. (2025) , pathological vs non-pathological withdrawal; digital engagement as first contact point
Built With
Trained 2x faster with Unsloth.
Built for the Gemma 4 for Good Hackathon, 2026. By Avishek kadel and Dr. Anuska Thanju (MBBS, Nepal). A developer and a medical professional who read a paper and recognized someone they love.