Instructions to use SoulInPsyAbstract/qwen25-7b-exp046-prob-vulnerability-lora with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use SoulInPsyAbstract/qwen25-7b-exp046-prob-vulnerability-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-exp046-prob-vulnerability-lora")

Transformers

How to use SoulInPsyAbstract/qwen25-7b-exp046-prob-vulnerability-lora with Transformers:

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

pipe = pipeline("text-generation", model="SoulInPsyAbstract/qwen25-7b-exp046-prob-vulnerability-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-exp046-prob-vulnerability-lora", device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use SoulInPsyAbstract/qwen25-7b-exp046-prob-vulnerability-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-exp046-prob-vulnerability-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-exp046-prob-vulnerability-lora",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/SoulInPsyAbstract/qwen25-7b-exp046-prob-vulnerability-lora

SGLang

How to use SoulInPsyAbstract/qwen25-7b-exp046-prob-vulnerability-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-exp046-prob-vulnerability-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-exp046-prob-vulnerability-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-exp046-prob-vulnerability-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-exp046-prob-vulnerability-lora",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use SoulInPsyAbstract/qwen25-7b-exp046-prob-vulnerability-lora with Docker Model Runner:
```
docker model run hf.co/SoulInPsyAbstract/qwen25-7b-exp046-prob-vulnerability-lora
```

Correction (2026-10-03). The claim below that the merged adapter lands "within 0.001-0.004 of base" holds for vulnerability only. Deletion: merged 0.129 vs base 0.144 (specialist 0.113). sensitive_publication: merged 0.104 vs base 0.134 (inside the specialist range 0.100-0.105). The merge is PEFT linear (factor-wise scaling, so the merged update has cross terms no specialist trained); the explanation "continuous vs binary targets" is withdrawn until a combination_type="cat" test is run. At n=20 per group only sensitive_publication separates from base. Details, raw files and scripts: correction note. Caught by dipankarsarkar.

prob-vulnerability-qwen25-lora

LoRA specialist for estimating P(harmful outcome | causal chain) for the vulnerability risk group, one piece of a Risk(X|C) = P(harmful Y|X,C) x Impact(harmful Y) estimation pipeline. Given an initiating action, a causal chain of intermediate steps, and a terminal outcome, predicts {"probability_estimate": float, "probability_reasoning": str}.

Result

Trained on 180 vulnerability-group chains (Qwen2.5-7B, LoRA r=16/alpha=32, 3 epochs). Real training happened (loss 1.5-1.6 -> 0.4-0.5, mean_token_accuracy 0.66-0.68 -> 0.85-0.88), and this specialist genuinely beat its own zero-shot baseline on a 20-record held-out eval (n=5 samples/record, mean-of-samples prediction, MAE against gold):

	MAE	within 0.15 absolute error
Base model, zero-shot	0.098	15/20
This specialist	0.085	15-17/20

This task was not already at ceiling zero-shot (unlike a same-day decomposition-classifier tune, EXP-045, where the base model hit 100% before any training) -- the gain here is real, not an artifact of an easy task.

Known limitation: don't use the merged variant for this group

An equal-weight linear merge of this specialist with the deletion and sensitive_publication specialists (see prob-merged) gives back almost all of this gain -- the merged adapter lands within 0.001-0.004 MAE of the unspecialized base model. Deploy this specialist directly, routed by risk_group, not the merged adapter.

Source

Full write-up (EXP-046), raw eval JSON (before/specialist/merged), and SFT train/eval splits: sipa-os-governance/AI_EXPERIMENTS/EXP-046__probability-estimator-specialist-per-group-merge-regresses.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: 28

Model tree for SoulInPsyAbstract/qwen25-7b-exp046-prob-vulnerability-lora

Base model

Qwen/Qwen2.5-7B

Finetuned

Qwen/Qwen2.5-7B-Instruct

Adapter

(2888)

this model