H4L0

H4L0 is a QLoRA adapter on top of Qwen/Qwen3-8B.

It is a dual-mode assistant:

1. VRChat desktop companion - knows VRChat cold: keyboard/mouse controls, menus, safety settings, trust ranks, worlds, avatars, groups, instances. Its social style is snappy and human: one or two short lines, casual, warm, never a wall of text.
2. Security / engineering mode - the original H4L0 core: concise, direct technical answers across web development docs, web security and payload techniques, people skills, with a hard refuse on unauthorized real-world targets.

Training

• Method: QLoRA (4-bit NF4, bfloat16 compute), Unsloth + TRL SFTTrainer
• LoRA: r=16, alpha=32, no dropout, all linear projections
• Sequence length 1024, 1 epoch, lr 2e-4, batch 1 x grad-accum 8
• 240 optimizer steps, ~12.8 min on a single 12 GB RTX 5070
• Final train loss: ~1.44
• 1,919 examples (VRChat: 418; the rest: dev/docs, web security, people skills, persona)

Data

• VRChat Wiki and VRChat Help Center articles (controls, menus, safety, trust ranks, avatars, worlds, groups)
• Hand-written VRChat social dialogue set (greetings, callouts, conflict de-escalation, boundaries, new-player help)
• Public developer documentation (devdocs.io), PortSwigger Web Security Academy, PayloadsAllTheThings (webshell/exploit payloads scrubbed), Wikipedia soft-skills articles (CC BY-SA)
• Persona and refusal sets

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
import torch

base = "Qwen/Qwen3-8B"
adapter = "sauceyy909/H4l0"

tok = AutoTokenizer.from_pretrained(base)
model = AutoModelForCausalLM.from_pretrained(
    base, torch_dtype=torch.bfloat16, device_map="auto")
model = PeftModel.from_pretrained(model, adapter)
model.eval()

messages = [{"role": "user", "content": "yo, can I join you guys?"}]
ids = tok.apply_chat_template(
    messages, add_generation_prompt=True, return_tensors="pt",
    enable_thinking=False).to(model.device)
out = model.generate(ids, max_new_tokens=64, temperature=0.7, top_p=0.9,
                     do_sample=True, pad_token_id=tok.pad_token_id)
print(tok.decode(out[0][ids.shape[1]:], skip_special_tokens=True))

For VRChat-style replies, keep max_new_tokens short (40-80).

Intended use and limits

Personal companion / coaching assistant. It is not a game cheat, not a moderation tool, and not an authority on VRChat policy. Security content is for authorized testing, labs, CTFs and your own systems only. The model can be wrong - verify anything that matters. H4L0 is a fan-made persona and is not affiliated with VRChat Inc.