Instructions to use youngseok12/AX-3.1-Light-specialist-r1-ties with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use youngseok12/AX-3.1-Light-specialist-r1-ties with Transformers:

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

pipe = pipeline("text-generation", model="youngseok12/AX-3.1-Light-specialist-r1-ties")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("youngseok12/AX-3.1-Light-specialist-r1-ties")
model = AutoModelForCausalLM.from_pretrained("youngseok12/AX-3.1-Light-specialist-r1-ties", device_map="auto")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=40)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use youngseok12/AX-3.1-Light-specialist-r1-ties with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "youngseok12/AX-3.1-Light-specialist-r1-ties"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "youngseok12/AX-3.1-Light-specialist-r1-ties",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/youngseok12/AX-3.1-Light-specialist-r1-ties

SGLang

How to use youngseok12/AX-3.1-Light-specialist-r1-ties 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 "youngseok12/AX-3.1-Light-specialist-r1-ties" \
    --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": "youngseok12/AX-3.1-Light-specialist-r1-ties",
		"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 "youngseok12/AX-3.1-Light-specialist-r1-ties" \
        --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": "youngseok12/AX-3.1-Light-specialist-r1-ties",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use youngseok12/AX-3.1-Light-specialist-r1-ties with Docker Model Runner:
```
docker model run hf.co/youngseok12/AX-3.1-Light-specialist-r1-ties
```

A.X-3.1-Light — Specialist LoRA Merge (`ties`)

세 개의 축별 specialist LoRA(K/R/C)를 각각 따로 학습한 뒤 TIES로 병합한 BF16 전체 가중치 모델입니다. 별도 adapter가 필요하지 않습니다.

연구·평가용 모델입니다. 생성 결과가 부정확할 수 있으므로 의료·법률·재정 등 고위험 의사결정의 유일한 근거로 사용해서는 안 됩니다.

Model information

Base model: skt/A.X-3.1-Light
Base revision: 9b41bb2406472634d8812c0b8931fa40fa9a6c3a
Merge method: TIES — K + R + C, weights 1.0 each, density 0.5
Format: standalone BF16 safetensors
Chat template: base 원본 그대로 (수정하지 않음)

Specialist LoRAs

각 LoRA는 담당 축의 데이터만 3,000행 학습했고, 학습 조건은 세 개가 완전히 동일합니다 (유일한 변수는 데이터). 전 축 4지선다 ①②③④ + answer-first 한 줄 타깃 스키마.

LoRA	Target axis	Dataset	Source
K-LoRA	KMMLU-Pro	AI Hub 71875 필수의료 의학지식	link
R-LoRA	MuSR(Ko)	AI Hub 71568 숫자연산 기계독해	link
C-LoRA	Com2-main(Ko)	AI Hub 71949 인과관계 기반 추론	link

Train loss: K 0.1302 / R 0.1187 / C 0.0334
Held-out format probe (240 canonical public-benchmark items, greedy): answer stated in 100.0% of outputs, mean 13 output tokens

전처리

5지선다 → 4지선다 변환 (공식 벤치마크가 4지선다)
정답 위치 균등화 — 세 LoRA 모두 ①②③④ 각 750개
71568: calculation 수식이 있는 정답만 사용, 오답을 수식 변형으로 생성(검증 가능)
71949: 원본이 이미지 기반 MCQ이므로 텍스트 인과 CoT를 시드로 4지선다 합성, 이미지를 지칭하는 항목은 전량 제외

Training configuration (three LoRAs identical)


Method	LoRA, merged for inference
Rank / alpha / dropout	16 / 32 / 0.05
Target modules	`q_proj`, `k_proj`, `v_proj`, `o_proj`, `gate_proj`, `up_proj`, `down_proj`
Learning rate	`5e-5`, cosine, warmup ratio 0.03
Epochs	1
Effective batch	16
Max sequence length	4,096
Precision	BF16
Seed	42
Objective	assistant-token-only causal LM cross entropy

Notes

공개 벤치마크 test set을 직접 학습하지 않았습니다.
병합은 단일 GPU에서 수행했고, 전 파라미터 NaN/Inf 검사를 통과했습니다.

Downloads last month: 300

Safetensors

Model size

7B params

Tensor type

BF16

Model tree for youngseok12/AX-3.1-Light-specialist-r1-ties

Base model

skt/A.X-3.1-Light

Adapter

(16)

this model

A.X-3.1-Light — Specialist LoRA Merge (ties)