Instructions to use werty1248/LRCLFM2MoE-base-checkpoint-160000 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use werty1248/LRCLFM2MoE-base-checkpoint-160000 with Transformers:

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

pipe = pipeline("text-generation", model="werty1248/LRCLFM2MoE-base-checkpoint-160000", trust_remote_code=True)
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("werty1248/LRCLFM2MoE-base-checkpoint-160000", trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained("werty1248/LRCLFM2MoE-base-checkpoint-160000", trust_remote_code=True, 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 werty1248/LRCLFM2MoE-base-checkpoint-160000 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "werty1248/LRCLFM2MoE-base-checkpoint-160000"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "werty1248/LRCLFM2MoE-base-checkpoint-160000",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/werty1248/LRCLFM2MoE-base-checkpoint-160000

SGLang

How to use werty1248/LRCLFM2MoE-base-checkpoint-160000 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 "werty1248/LRCLFM2MoE-base-checkpoint-160000" \
    --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": "werty1248/LRCLFM2MoE-base-checkpoint-160000",
		"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 "werty1248/LRCLFM2MoE-base-checkpoint-160000" \
        --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": "werty1248/LRCLFM2MoE-base-checkpoint-160000",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use werty1248/LRCLFM2MoE-base-checkpoint-160000 with Docker Model Runner:
```
docker model run hf.co/werty1248/LRCLFM2MoE-base-checkpoint-160000
```

Original Model: LiquidAI/LFM2-8B-A1B

Base 모델입니다.

REAP으로 6B-A1B로 변경
- 사용 데이터: allenai/tulu-3-sft-personas-math
- 다국어 능력/코딩 능력을 제거하고 나머지 능력을 남기기 위해 다양한 데이터 조합을 실험했지만, 수학 계열 데이터를 쓰는 게 제일 보존되는 평균 점수가 높았음.
LowRankClone으로 3B-A0.6B로 변경
- 사용 데이터: 약 10B 토큰
  - 70% HuggingFaceTB/smollm-corpus/fineweb-edu-dedup (score 3.5 이상)
  - 25% C10X/finepdfs-edu-hq
  - 5% OpenHermes-2.5
- 원래 논문에서는 Fineweb-edu + OpenHermes 사용

5-shot 벤치(lm-evaluation-harness, loglikelihood test)

ARC-C, HellaSwag는 acc_norm 사용

Name	Param	Active	MMLU	GPQA_main	PIQA	ARC-C	HellaSwag
LiquidAI/LFM2-8B-A1B	8.3B	1.2B	64.84	25.89	76.44	60.75	73.34
LiquidAI/LFM2-2.6B	2.6B	2.6B	64.62	35.04	77.53	56.14	72.37
LiquidAI/LFM2-1.2B	1.2B	1.2B	55.12	30.36	73.50	56.06	63.46
LiquidAI/LFM2-700M	0.7B	0.7B	49.42	31.03	72.25	51.19	58.02
werty1248/LFM2-6B-A1B-REAP	6.3B	1.2B	56.43	32.37	71.27	52.73	65.33
werty1248/LRCLFM2MoE-base-checkpoint-160000	3.2B	0.6B	44.49	29.69	71.11	49.40	54.77

교훈: REAP는 모든 학습이 끝난 다음에 하자(distillation이 나쁘게 됨)

총 비용: 2x B200 에서 110시간 (~$600)

Downloads last month: 20

Safetensors

Model size

3B params

Tensor type

F32

BF16