My-Jarvis 72B

A bilingual (Japanese / English) personal assistant model fine-tuned from Qwen2.5-72B-Instruct — the largest dense open model that can be QLoRA-trained on a single HAKUSAN GPU node (2 x A40 48GB). The weights contain no owner or user personal data: personalization starts only after each user loads the model, through onboarding questions, local conversation memory, and (optionally) a user-specific LoRA adapter trained on the user's own machine.

Model overview

Training data (public only, zero personal data)

• Aya Dataset (CohereLabs/aya_dataset): Japanese and English instruction rows (Apache-2.0).
• Japanese instruction corpus (izumi-lab/llm-japanese-dataset): capped sample (CC BY-SA 4.0).
• Persona seed: none. PERSONA_SEED is empty by design, so the published weights carry no identity. The model is instructed to learn the user after they load it.

No owner-curated identity corpus is used. A fail-closed PII scanner runs on the upload folder before publishing and aborts the upload on any hit (emails, phone-like strings, tokens, or an owner blocklist).

Personalization starts after load (2026 best practices)

The released weights are a blank slate; personalization lives entirely on the user's device:

1. Onboarding — on first run the assistant asks 3 questions (name, interests, tone) and saves them to ~/.my-jarvis/profile.json.
2. Proactive memory (ProMem-style) — after each turn the model distills new user facts into ~/.my-jarvis/memory.jsonl; retrieved memories are re-injected into the next prompt.
3. Per-user adapter (PLUME/Aplaud-style, optional) — export_lora_rows() converts local memory into a LoRA dataset so each user can train their own small adapter locally; it never leaves their machine.

from my_jarvis_pipeline import load_profile, load_memory, build_messages

profile = load_profile("~/.my-jarvis/profile.json")
memory = load_memory("~/.my-jarvis/memory.jsonl", limit=6)
messages = build_messages(profile, memory, "今日の調子はどう？")

Use

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "LoNebula/my-jarvis-72b"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id, torch_dtype=torch.bfloat16, device_map="auto")

messages = [{"role": "system", "content": "You are My-Jarvis. Learn the user's style and answer in their language."},
            {"role": "user", "content": "難しいバグにぶつかったら？"}]
inputs = tokenizer.apply_chat_template(messages, add_generation_prompt=True,
                                       return_tensors="pt").to(model.device)
print(tokenizer.decode(model.generate(inputs, max_new_tokens=512)[0][inputs.shape[-1]:],
                       skip_special_tokens=True))

72B bf16 needs ~150GB of RAM/VRAM to load; use 4-bit loading (bitsandbytes) or multiple GPUs on smaller machines, or the Inference API.

Limitations

• Experimental, not benchmarked. It may be inaccurate or reproduce training text.
• It starts without any user's personal knowledge; personalization happens per user at load time.
• Training is capped (single-node 24h-class jobs), so this is a general bilingual assistant, not a task specialist.

License

Apache-2.0, subject to the base model license (Qwen2.5-72B-Instruct) and the training data licenses (Aya: Apache-2.0; izumi-lab: CC BY-SA 4.0).