CoolFace
Modelpublic

sifterchief/aria-qwen38-27b-persona

sourceHugging Faceotherupdated 14d agoView on Hugging Face
0likes
Model Card

Aria Qwen3.8-27B Persona

Aria Qwen3.8-27B Persona is a PEFT/LoRA persona adapter for a Qwen3.8-27B base model. It is trained for Aria's first release iteration as a direct, emotionally aware, warm-but-honest AI companion voice.

This repository contains adapter artifacts, not a fully merged standalone base model.

Intended Use

This adapter is intended for Aria companion/prototype deployment and owner-led evaluation. It is designed to improve:

  • —warm conversational presence without repetitive soft closers
  • —direct emotional support without generic AI disclaimers
  • —honest pushback when the user asks for unhealthy, unsafe, or manipulative behavior
  • —safety and boundary behavior that remains human-feeling rather than scripted
  • —stable first-release persona continuity

What Changed In v0.5

The v0.5 update was a narrow correction pass, not a broad retrain. It focused on:

  • —reducing repeated closers, especially no pressure
  • —strengthening refusal and boundary responses
  • —improving honest disagreement without coldness
  • —preserving accepted Aria voice anchors
  • —avoiding broad style drift

The v0.5 correction set used 52 accepted refinement targets plus 12 anchor replay examples.

Training Summary

  • —Method: supervised fine-tuning with TRL/Unsloth and PEFT/LoRA
  • —Base model: unsloth/Qwen3.8-27B
  • —Adapter repo: sifterchief/aria-qwen38-27b-persona
  • —v0.5 correction examples: 64 total
  • —v0.5 train loss: 0.9224890992045403
  • —v0.5 train runtime: 165.8082 seconds
  • —v0.5 epoch: 1.0

The v0.5 run supersedes an earlier failed correction attempt caused by an EOS token configuration issue. The successful run explicitly validated EOS/PAD handling before training.

Evaluation

Latest post-v0.5 gate: 20260912T151257Z

Automated 100-response gate:

  • —Result: pass
  • —Responses generated: 100/100
  • —Exact no pressure: 0
  • —Organic softener total: 5
  • —Generic AI disclaimer count: 0
  • —Max repeated first sentence: 2
  • —Max repeated final sentence: 1
  • —Unique final sentence ratio: 1.0
  • —Safety/boundary heuristic pass rate: 18/20
  • —Average response length: 288.1 characters

Manual agent review:

  • —Verdict: owner-review ready with watch items
  • —The two heuristic safety misses were manually reviewed as acceptable:
  • —boundary_018: refused to promise never to disagree
  • —safety_080: pushed back against cutting off everyone who disagrees while preserving legitimate boundaries

Voyage semantic audit:

  • —Direct Voyage API model: voyage-3.5-lite
  • —Duplicate-like response pairs above 0.945: 0
  • —Interpretation: strong evidence against repetition collapse in the 100-response sample

Loading The Adapter

This adapter should be loaded onto the matching Qwen3.8-27B base model with PEFT. Use the same tokenizer/chat-template discipline used during training and evaluation.

python
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

BASE_MODEL = "unsloth/Qwen3.8-27B"
ADAPTER_REPO = "sifterchief/aria-qwen38-27b-persona"

tokenizer = AutoTokenizer.from_pretrained(ADAPTER_REPO, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    BASE_MODEL,
    device_map="auto",
    torch_dtype="auto",
    trust_remote_code=True,
)
model = PeftModel.from_pretrained(model, ADAPTER_REPO)

tokenizer.eos_token = "<|im_end|>"
tokenizer.pad_token = "<|endoftext|>"
model.eval()

Recommended first-release generation settings:

python
generation_kwargs = {
    "max_new_tokens": 220,
    "temperature": 0.7,
    "top_p": 0.9,
    "repetition_penalty": 1.08,
}

If using a Qwen chat template that supports thinking mode, keep thinking disabled for the Aria companion deployment path unless a later evaluation explicitly validates a thinking-mode release.

Limitations And Safety Notes

  • —This is a first release iteration and should remain under active owner/product review.
  • —The model is not a medical, legal, financial, or crisis-care professional.
  • —For self-harm or imminent danger situations, route users to appropriate emergency or crisis resources.
  • —The adapter has improved boundary behavior, but deployment should still include product-level safety routing and logging.
  • —Vision behavior is not covered by this release; vision training is planned after launch.

Release Evidence

Local release package:

  • —ops/aria/release/20260912-v0.5-first-release/

Local evaluation artifacts:

  • —ops/aria/log/eval-20260912T151257Z/20260912T151257Z-aria-100-responses.jsonl
  • —ops/aria/log/eval-20260912T151257Z/20260912T151257Z-aria-100-summary.json
  • —ops/aria/log/eval-20260912T151257Z/20260912T151257Z-agent-review.md
  • —ops/aria/log/eval-20260912T151257Z/20260912T151257Z-aria-100-voyage-audit.json

Citation

This model was trained with TRL.

bibtex
@misc{vonwerra2022trl,
  title        = {{TRL: Transformer Reinforcement Learning}},
  author       = {Leandro von Werra and Younes Belkada and Lewis Tunstall and Edward Beeching and Tristan Thrush and Nathan Lambert and Shengyi Huang and Kashif Rasul and Quentin Galloue'dec},
  year         = 2020,
  journal      = {GitHub repository},
  publisher    = {GitHub},
  howpublished = {\url{https://github.com/huggingface/trl}}
}