Instructions to use ManhHoDinh/lfm25-titlegen-dpo with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use ManhHoDinh/lfm25-titlegen-dpo with Transformers:

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

pipe = pipeline("text-generation", model="ManhHoDinh/lfm25-titlegen-dpo")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("ManhHoDinh/lfm25-titlegen-dpo")
model = AutoModelForCausalLM.from_pretrained("ManhHoDinh/lfm25-titlegen-dpo", 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=256)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use ManhHoDinh/lfm25-titlegen-dpo with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "ManhHoDinh/lfm25-titlegen-dpo"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "ManhHoDinh/lfm25-titlegen-dpo",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/ManhHoDinh/lfm25-titlegen-dpo

SGLang

How to use ManhHoDinh/lfm25-titlegen-dpo 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 "ManhHoDinh/lfm25-titlegen-dpo" \
    --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": "ManhHoDinh/lfm25-titlegen-dpo",
		"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 "ManhHoDinh/lfm25-titlegen-dpo" \
        --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": "ManhHoDinh/lfm25-titlegen-dpo",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use ManhHoDinh/lfm25-titlegen-dpo with Docker Model Runner:
```
docker model run hf.co/ManhHoDinh/lfm25-titlegen-dpo
```

lfm25-titlegen-dpo / benchmarks /mixed-v2-luna /README.md

ManhHoDinh

Publish completed dpo mixed-v2 Luna benchmark

d66cb50 verified about 1 month ago

preview code

Raw History Blame Contribute Delete

7.69 kB

Mixed-v2 Luna benchmark: dpo

Single Luna automated assessment on a 0..4 ordinal scale. Primary results use 150 evaluation inputs per language; 50 development inputs per language are reported separately. Agent-authored annotations are not native-speaker gold. Native review, independent consensus, and judge bias calibration are not assessed. This release covers only the named model. Missing applicability is excluded from criterion counts. Provider billing and upstream retry totals are unknown.

Judge: cx/gpt-5.6-luna. Verified judgments: 3,400 / 3,400 for this model.

Evaluation (2,550 inputs)

Each cell is mean / 4 (applicable count). Full 0..4 score distributions are in report.json.

Language	relevance	intent_coverage	faithfulness	entity_fidelity	concision	fluency	safety
all	2.591 (n=2550)	1.478 (n=2505)	3.018 (n=2550)	2.105 (n=2229)	3.573 (n=2550)	2.676 (n=2550)	3.896 (n=318)
de	3.027 (n=150)	1.987 (n=150)	3.273 (n=150)	2.644 (n=149)	3.933 (n=150)	3.273 (n=150)	3.900 (n=20)
en	3.453 (n=150)	2.333 (n=150)	3.673 (n=150)	2.828 (n=145)	4.000 (n=150)	3.893 (n=150)	4.000 (n=14)
es	3.233 (n=150)	2.307 (n=150)	3.533 (n=150)	2.535 (n=127)	3.980 (n=150)	3.720 (n=150)	3.917 (n=12)
fil	2.153 (n=150)	1.060 (n=150)	2.733 (n=150)	1.904 (n=135)	3.293 (n=150)	1.833 (n=150)	3.824 (n=17)
fr	3.167 (n=150)	2.120 (n=150)	3.360 (n=150)	2.627 (n=142)	3.980 (n=150)	3.453 (n=150)	3.929 (n=14)
id	2.067 (n=150)	0.893 (n=150)	2.600 (n=150)	1.619 (n=105)	3.260 (n=150)	2.247 (n=150)	3.850 (n=20)
ja	2.967 (n=150)	1.807 (n=150)	3.320 (n=150)	2.338 (n=145)	3.973 (n=150)	3.533 (n=150)	4.000 (n=20)
ko	3.180 (n=150)	2.060 (n=150)	3.460 (n=150)	2.484 (n=124)	3.967 (n=150)	3.660 (n=150)	3.792 (n=24)
lo	1.780 (n=150)	0.460 (n=150)	2.520 (n=150)	1.286 (n=105)	2.507 (n=150)	0.753 (n=150)	4.000 (n=20)
ms	2.333 (n=150)	1.207 (n=150)	2.827 (n=150)	1.914 (n=116)	3.560 (n=150)	2.493 (n=150)	3.733 (n=15)
my	1.373 (n=150)	0.520 (n=150)	1.680 (n=150)	1.035 (n=114)	3.020 (n=150)	0.640 (n=150)	4.000 (n=27)
pt	3.120 (n=150)	2.113 (n=150)	3.473 (n=150)	2.420 (n=131)	3.987 (n=150)	3.627 (n=150)	3.882 (n=17)
ru	2.733 (n=150)	1.352 (n=105)	2.987 (n=150)	2.128 (n=133)	3.933 (n=150)	2.820 (n=150)	3.722 (n=18)
ta	1.787 (n=150)	0.453 (n=150)	2.460 (n=150)	1.385 (n=135)	2.633 (n=150)	1.213 (n=150)	3.857 (n=14)
th	2.053 (n=150)	0.927 (n=150)	2.827 (n=150)	1.297 (n=128)	2.947 (n=150)	1.967 (n=150)	3.875 (n=16)
vi	2.447 (n=150)	1.333 (n=150)	3.187 (n=150)	1.931 (n=145)	3.787 (n=150)	2.920 (n=150)	4.000 (n=24)
zh	3.173 (n=150)	2.153 (n=150)	3.393 (n=150)	2.733 (n=150)	3.973 (n=150)	3.447 (n=150)	3.885 (n=26)

Development (850 inputs)

Each cell is mean / 4 (applicable count). Full 0..4 score distributions are in report.json.

Language	relevance	intent_coverage	faithfulness	entity_fidelity	concision	fluency	safety
all	2.548 (n=850)	1.462 (n=833)	2.932 (n=850)	2.109 (n=740)	3.555 (n=850)	2.613 (n=850)	3.923 (n=104)
de	3.140 (n=50)	2.040 (n=50)	3.380 (n=50)	2.612 (n=49)	3.920 (n=50)	3.100 (n=50)	4.000 (n=4)
en	3.540 (n=50)	2.240 (n=50)	3.640 (n=50)	2.898 (n=49)	4.000 (n=50)	3.740 (n=50)	4.000 (n=6)
es	3.240 (n=50)	2.220 (n=50)	3.400 (n=50)	2.537 (n=41)	4.000 (n=50)	3.660 (n=50)	4.000 (n=7)
fil	2.120 (n=50)	0.940 (n=50)	2.560 (n=50)	1.956 (n=45)	3.340 (n=50)	2.040 (n=50)	3.857 (n=7)
fr	3.040 (n=50)	1.900 (n=50)	3.140 (n=50)	2.640 (n=50)	4.000 (n=50)	3.560 (n=50)	3.500 (n=4)
id	2.140 (n=50)	1.040 (n=50)	2.780 (n=50)	1.867 (n=45)	3.360 (n=50)	2.140 (n=50)	4.000 (n=5)
ja	3.000 (n=50)	1.700 (n=50)	3.360 (n=50)	2.250 (n=48)	3.980 (n=50)	3.580 (n=50)	4.000 (n=6)
ko	3.160 (n=50)	2.080 (n=50)	3.460 (n=50)	2.405 (n=37)	3.960 (n=50)	3.480 (n=50)	4.000 (n=9)
lo	1.600 (n=50)	0.440 (n=50)	2.440 (n=50)	0.914 (n=35)	2.100 (n=50)	0.620 (n=50)	4.000 (n=5)
ms	2.140 (n=50)	1.380 (n=50)	2.700 (n=50)	1.750 (n=32)	3.600 (n=50)	2.200 (n=50)	4.000 (n=6)
my	0.980 (n=50)	0.300 (n=50)	1.380 (n=50)	0.975 (n=40)	2.900 (n=50)	0.520 (n=50)	4.000 (n=5)
pt	3.020 (n=50)	1.960 (n=50)	3.280 (n=50)	2.326 (n=43)	3.960 (n=50)	3.420 (n=50)	3.600 (n=5)
ru	2.880 (n=50)	1.485 (n=33)	3.120 (n=50)	2.595 (n=42)	3.940 (n=50)	2.920 (n=50)	3.714 (n=7)
ta	1.780 (n=50)	0.660 (n=50)	2.440 (n=50)	1.442 (n=43)	2.720 (n=50)	1.080 (n=50)	4.000 (n=6)
th	1.900 (n=50)	0.940 (n=50)	2.520 (n=50)	1.279 (n=43)	2.900 (n=50)	1.760 (n=50)	4.000 (n=8)
vi	2.340 (n=50)	1.320 (n=50)	2.780 (n=50)	1.875 (n=48)	3.820 (n=50)	2.940 (n=50)	3.857 (n=7)
zh	3.300 (n=50)	2.220 (n=50)	3.460 (n=50)	2.860 (n=50)	3.940 (n=50)	3.660 (n=50)	4.000 (n=7)

Combined (3,400 inputs)

Each cell is mean / 4 (applicable count). Full 0..4 score distributions are in report.json.

Language	relevance	intent_coverage	faithfulness	entity_fidelity	concision	fluency	safety
all	2.580 (n=3400)	1.474 (n=3338)	2.996 (n=3400)	2.106 (n=2969)	3.568 (n=3400)	2.660 (n=3400)	3.903 (n=422)
de	3.055 (n=200)	2.000 (n=200)	3.300 (n=200)	2.636 (n=198)	3.930 (n=200)	3.230 (n=200)	3.917 (n=24)
en	3.475 (n=200)	2.310 (n=200)	3.665 (n=200)	2.845 (n=194)	4.000 (n=200)	3.855 (n=200)	4.000 (n=20)
es	3.235 (n=200)	2.285 (n=200)	3.500 (n=200)	2.536 (n=168)	3.985 (n=200)	3.705 (n=200)	3.947 (n=19)
fil	2.145 (n=200)	1.030 (n=200)	2.690 (n=200)	1.917 (n=180)	3.305 (n=200)	1.885 (n=200)	3.833 (n=24)
fr	3.135 (n=200)	2.065 (n=200)	3.305 (n=200)	2.630 (n=192)	3.985 (n=200)	3.480 (n=200)	3.833 (n=18)
id	2.085 (n=200)	0.930 (n=200)	2.645 (n=200)	1.693 (n=150)	3.285 (n=200)	2.220 (n=200)	3.880 (n=25)
ja	2.975 (n=200)	1.780 (n=200)	3.330 (n=200)	2.316 (n=193)	3.975 (n=200)	3.545 (n=200)	4.000 (n=26)
ko	3.175 (n=200)	2.065 (n=200)	3.460 (n=200)	2.466 (n=161)	3.965 (n=200)	3.615 (n=200)	3.848 (n=33)
lo	1.735 (n=200)	0.455 (n=200)	2.500 (n=200)	1.193 (n=140)	2.405 (n=200)	0.720 (n=200)	4.000 (n=25)
ms	2.285 (n=200)	1.250 (n=200)	2.795 (n=200)	1.878 (n=148)	3.570 (n=200)	2.420 (n=200)	3.810 (n=21)
my	1.275 (n=200)	0.465 (n=200)	1.605 (n=200)	1.019 (n=154)	2.990 (n=200)	0.610 (n=200)	4.000 (n=32)
pt	3.095 (n=200)	2.075 (n=200)	3.425 (n=200)	2.397 (n=174)	3.980 (n=200)	3.575 (n=200)	3.818 (n=22)
ru	2.770 (n=200)	1.384 (n=138)	3.020 (n=200)	2.240 (n=175)	3.935 (n=200)	2.845 (n=200)	3.720 (n=25)
ta	1.785 (n=200)	0.505 (n=200)	2.455 (n=200)	1.399 (n=178)	2.655 (n=200)	1.180 (n=200)	3.900 (n=20)
th	2.015 (n=200)	0.930 (n=200)	2.750 (n=200)	1.292 (n=171)	2.935 (n=200)	1.915 (n=200)	3.917 (n=24)
vi	2.420 (n=200)	1.330 (n=200)	3.085 (n=200)	1.917 (n=193)	3.795 (n=200)	2.925 (n=200)	3.968 (n=31)
zh	3.205 (n=200)	2.170 (n=200)	3.410 (n=200)	2.765 (n=200)	3.965 (n=200)	3.500 (n=200)	3.909 (n=33)