---
base_model: Jackrong/Qwopus3.6-35B-A3B-v1
library_name: peft
pipeline_tag: image-text-to-text
license: apache-2.0
tags:
- qwen3_5_moe
- qwen3.6
- qwopus
- moe
- lora
- peft
- unsloth
- union-street-ai
- helios
- identity-tune
- adapter
language:
- en
---

# Helios Rabbit v1

Helios Rabbit v1 is a lightweight identity and behavior LoRA adapter for `Jackrong/Qwopus3.6-35B-A3B-v1`, produced by Union Street AI.

This is an adapter, not a full merged checkpoint. Use it with the base model named above.

## Intended Identity

The adapter is intended to make the model identify as Helios, a local AI model developed and adapted by Union Street AI, while preserving the base model's coding, repo-analysis, and infrastructure strengths.

It should be honest about lineage: Helios is adapted from open model research and local post-training work. It should not claim that Union Street AI trained the base model from scratch.

## Training Summary

- Run name: `helios-rabbit-v1`
- Base model: `Jackrong/Qwopus3.6-35B-A3B-v1`
- Method: LoRA SFT with Unsloth / PEFT
- Data: 475 training conversations, 25 validation conversations
- Max sequence length: 2048
- LoRA rank: 8
- LoRA alpha: 8
- Target: language attention modules, vision layers disabled, MLP expert LoRA disabled for this first identity pass
- Hardware: Lambda Labs 8x NVIDIA A100-SXM4-80GB

## Dataset Notes

The dataset is a small synthetic identity and behavior corpus for Helios. It focuses on:

- identity and provenance
- coding and infrastructure assistant behavior
- candid but bounded adult-world conversation
- liberty-minded, anti-authoritarian, rule-of-law, pro-human-agency posture
- honesty about uncertainty and model lineage

## Status

This is a v1 experimental adapter. Evaluate before production use.

## Loading Sketch

```python
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_id = "Jackrong/Qwopus3.6-35B-A3B-v1"
adapter_id = "UnionStreet/helios-rabbit-v1"

tokenizer = AutoTokenizer.from_pretrained(base_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(base_id, torch_dtype="auto", device_map="auto", trust_remote_code=True)
model = PeftModel.from_pretrained(model, adapter_id)
```

Depending on your inference stack, you may need the multimodal Qwen3.5 MoE model class rather than `AutoModelForCausalLM`.

## License

The base model card declares `apache-2.0`; this adapter is released under Apache 2.0 as well, subject to the base model's terms.