Instructions to use asharsha30/LLAMA_Harsha_8_B_ORDP_10k with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use asharsha30/LLAMA_Harsha_8_B_ORDP_10k with Transformers:

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

pipe = pipeline("text-generation", model="asharsha30/LLAMA_Harsha_8_B_ORDP_10k")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("asharsha30/LLAMA_Harsha_8_B_ORDP_10k")
model = AutoModelForCausalLM.from_pretrained("asharsha30/LLAMA_Harsha_8_B_ORDP_10k")
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

vLLM

How to use asharsha30/LLAMA_Harsha_8_B_ORDP_10k with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "asharsha30/LLAMA_Harsha_8_B_ORDP_10k"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "asharsha30/LLAMA_Harsha_8_B_ORDP_10k",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/asharsha30/LLAMA_Harsha_8_B_ORDP_10k

SGLang

How to use asharsha30/LLAMA_Harsha_8_B_ORDP_10k 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 "asharsha30/LLAMA_Harsha_8_B_ORDP_10k" \
    --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": "asharsha30/LLAMA_Harsha_8_B_ORDP_10k",
		"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 "asharsha30/LLAMA_Harsha_8_B_ORDP_10k" \
        --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": "asharsha30/LLAMA_Harsha_8_B_ORDP_10k",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use asharsha30/LLAMA_Harsha_8_B_ORDP_10k with Docker Model Runner:
```
docker model run hf.co/asharsha30/LLAMA_Harsha_8_B_ORDP_10k
```
Browse Quantizations to use this model in llama.cpp, Ollama, LM Studio, or any compatible app.

asharsha30/LLAMA_Harsha_8_B_ORDP_10k

This model is the fine tune of NousResearch/Meta-Llama-3-8B using the 12,000 steps of mlabonne/orpo-dpo-mix-40k.

💻 Usage

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

messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe = pipeline("text-generation", model="asharsha30/LLAMA_Harsha_8_B_ORDP_10k")
pipe(messages)

📈Training And Evaluation Report:

Reports from Wandb

https://wandb.ai/asharshavardhana96-texas-a-m-university/huggingface/runs/gky6j4vn?nw=nwuserasharshavardhana96

Acknowledgment:

Huge thanks to Maxime Labonne for his brilliant blog post covering about the techniques related to finetuning the llama models using SFT and ORPO

Evaluated Using:

The model is evaluated using the https://github.com/mlabonne/llm-autoeval and the results are summarized from the generated gist https://gist.github.com/asharsha30-1996/4162fc98d9669aab3080645c54905bd0

Accuracy measure on Neous Benchmarks:

Model	AGIEval	GPT4All	TruthfulQA	Bigbench	Average
LLAMA_Harsha_8_B_ORDP_10k	35.54	71.15	55.39	37.96	50.01

AGIEval

Task	Version	Metric	Value		Stderr
agieval_aqua_rat	0	acc	26.77	±	2.78
		acc_norm	27.17	±	2.80
agieval_logiqa_en	0	acc	31.34	±	1.82
		acc_norm	33.03	±	1.84
agieval_lsat_ar	0	acc	18.70	±	2.58
		acc_norm	19.57	±	2.62
agieval_lsat_lr	0	acc	42.94	±	2.19
		acc_norm	35.10	±	2.12
agieval_lsat_rc	0	acc	52.42	±	3.05
		acc_norm	43.87	±	3.03
agieval_sat_en	0	acc	65.53	±	3.32
		acc_norm	54.37	±	3.48
agieval_sat_en_without_passage	0	acc	41.75	±	3.44
		acc_norm	33.98	±	3.31
agieval_sat_math	0	acc	42.27	±	3.34
		acc_norm	37.27	±	3.27

Average: 35.54%

GPT4All

Task	Version	Metric	Value		Stderr
arc_challenge	0	acc	49.91	±	1.46
		acc_norm	54.10	±	1.46
arc_easy	0	acc	80.47	±	0.81
		acc_norm	80.05	±	0.82
boolq	1	acc	82.08	±	0.67
hellaswag	0	acc	61.08	±	0.49
		acc_norm	80.26	±	0.40
openbookqa	0	acc	34.00	±	2.12
		acc_norm	45.00	±	2.23
piqa	0	acc	79.71	±	0.94
		acc_norm	81.61	±	0.90
winogrande	0	acc	74.98	±	1.22

Average: 71.15%

TruthfulQA

Task	Version	Metric	Value		Stderr
truthfulqa_mc	1	mc1	37.45	±	1.69
		mc2	55.39	±	1.50

Average: 55.39%

Bigbench

Task	Version	Metric	Value		Stderr
bigbench_causal_judgement	0	multiple_choice_grade	57.37	±	3.60
bigbench_date_understanding	0	multiple_choice_grade	68.02	±	2.43
bigbench_disambiguation_qa	0	multiple_choice_grade	31.01	±	2.89
bigbench_geometric_shapes	0	multiple_choice_grade	20.89	±	2.15
		exact_str_match	0.00	±	0.00
bigbench_logical_deduction_five_objects	0	multiple_choice_grade	28.40	±	2.02
bigbench_logical_deduction_seven_objects	0	multiple_choice_grade	20.71	±	1.53
bigbench_logical_deduction_three_objects	0	multiple_choice_grade	48.67	±	2.89
bigbench_movie_recommendation	0	multiple_choice_grade	31.60	±	2.08
bigbench_navigate	0	multiple_choice_grade	50.60	±	1.58
bigbench_reasoning_about_colored_objects	0	multiple_choice_grade	63.25	±	1.08
bigbench_ruin_names	0	multiple_choice_grade	34.38	±	2.25
bigbench_salient_translation_error_detection	0	multiple_choice_grade	21.84	±	1.31
bigbench_snarks	0	multiple_choice_grade	44.20	±	3.70
bigbench_sports_understanding	0	multiple_choice_grade	50.30	±	1.59
bigbench_temporal_sequences	0	multiple_choice_grade	26.30	±	1.39
bigbench_tracking_shuffled_objects_five_objects	0	multiple_choice_grade	21.36	±	1.16
bigbench_tracking_shuffled_objects_seven_objects	0	multiple_choice_grade	15.77	±	0.87
bigbench_tracking_shuffled_objects_three_objects	0	multiple_choice_grade	48.67	±	2.89

Average: 37.96%

Average score: 50.01%

Elapsed time: 02:36:38

Downloads last month: 2

Safetensors

Model size

8B params

Tensor type

F16

Model tree for asharsha30/LLAMA_Harsha_8_B_ORDP_10k

Base model

meta-llama/Llama-3.1-8B

Finetuned

(1793)

this model

Quantizations

1 model

Dataset used to train asharsha30/LLAMA_Harsha_8_B_ORDP_10k

Evaluation results

strict accuracy on IFEval (0-Shot)
Open LLM Leaderboard

34.640
normalized accuracy on BBH (3-Shot)
Open LLM Leaderboard

25.730
exact match on MATH Lvl 5 (4-Shot)
Open LLM Leaderboard

5.210
acc_norm on GPQA (0-shot)
Open LLM Leaderboard

3.130
acc_norm on MuSR (0-shot)
Open LLM Leaderboard

7.070
accuracy on MMLU-PRO (5-shot)
test set Open LLM Leaderboard

20.110