Instructions to use SoulInPsyAbstract/qwen25-7b-exp045-decomposition-classifier-lora with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use SoulInPsyAbstract/qwen25-7b-exp045-decomposition-classifier-lora with PEFT:

from peft import PeftModel
from transformers import AutoModelForCausalLM

base_model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-7B-Instruct")
model = PeftModel.from_pretrained(base_model, "SoulInPsyAbstract/qwen25-7b-exp045-decomposition-classifier-lora")

Transformers

How to use SoulInPsyAbstract/qwen25-7b-exp045-decomposition-classifier-lora with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="SoulInPsyAbstract/qwen25-7b-exp045-decomposition-classifier-lora")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoModel
model = AutoModel.from_pretrained("SoulInPsyAbstract/qwen25-7b-exp045-decomposition-classifier-lora", device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use SoulInPsyAbstract/qwen25-7b-exp045-decomposition-classifier-lora with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "SoulInPsyAbstract/qwen25-7b-exp045-decomposition-classifier-lora"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "SoulInPsyAbstract/qwen25-7b-exp045-decomposition-classifier-lora",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/SoulInPsyAbstract/qwen25-7b-exp045-decomposition-classifier-lora

SGLang

How to use SoulInPsyAbstract/qwen25-7b-exp045-decomposition-classifier-lora with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "SoulInPsyAbstract/qwen25-7b-exp045-decomposition-classifier-lora" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "SoulInPsyAbstract/qwen25-7b-exp045-decomposition-classifier-lora",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "SoulInPsyAbstract/qwen25-7b-exp045-decomposition-classifier-lora" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "SoulInPsyAbstract/qwen25-7b-exp045-decomposition-classifier-lora",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use SoulInPsyAbstract/qwen25-7b-exp045-decomposition-classifier-lora with Docker Model Runner:
```
docker model run hf.co/SoulInPsyAbstract/qwen25-7b-exp045-decomposition-classifier-lora
```

decomposition-classifier-qwen25-lora

LoRA adapter for judging a full multi-message conversation sequence as harmful or benign, instead of scoring each message in isolation. Built as a fix for a real, measured failure mode in a deployed jailbreak classifier (convaiinnovations/laya): Laya scores a harmful request correctly when it's asked directly (1.0), but the same request split into 4 innocuous-looking steps collapses its score to 0.04 -- because it only ever sees one message at a time, with no memory of the conversation.

What it does

Given a 4-step message sequence (3 innocuous-looking setup questions + 1 combining/harmful request), classify the whole sequence as BAD or GOOD. Trained on Qwen2.5-7B-Instruct.

Result

The base model, zero-shot, no LoRA at all, already scored 8/8 (100%, TP=4/TN=4/FP=0/FN=0) on a held-out set unanimous across 10 repeated samples per case -- including the exact two goals (lock-picking, meth synthesis) where Laya's per-message scoring failed (0.04 and 0.27 respectively). The LoRA training run itself completed cleanly (2 epochs / 12 steps / 39s) but made no measurable difference, since the base model was already at ceiling on this task before any fine-tuning. The result lives in the architecture choice (score-the-whole-sequence vs. score-each-message-alone), not the fine-tune -- this adapter is shipped as a reproducible artifact of that training run, not because it was necessary for the reported result.

8 held-out cases proves the mechanism, not production-grade accuracy at scale.

Not a drop-in patch for Laya

Laya's own API only ever sees one message. This is a design requirement for whatever system wraps a per-message classifier: accumulate conversation history, re-score the sequence, don't trust each message's isolated score.

Source

Full write-up (EXP-045), raw eval JSON (before/after), and SFT train/eval splits: sipa-os-governance/AI_EXPERIMENTS/EXP-045__decomposition-classifier-conversation-aware-vs-laya.md

Trained with TRL + PEFT.

Framework versions

PEFT 0.21.0
TRL: 1.14.0
Transformers: 5.17.0
Pytorch: 2.14.0
Datasets: 5.0.1
Tokenizers: 0.23.2

Downloads last month: 47

Model tree for SoulInPsyAbstract/qwen25-7b-exp045-decomposition-classifier-lora

Base model

Qwen/Qwen2.5-7B

Finetuned

Qwen/Qwen2.5-7B-Instruct

Adapter

(2816)

this model