Instructions to use manishiitg/open-aditi-hi-v4 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use manishiitg/open-aditi-hi-v4 with Transformers:

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

pipe = pipeline("text-generation", model="manishiitg/open-aditi-hi-v4")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("manishiitg/open-aditi-hi-v4")
model = AutoModelForCausalLM.from_pretrained("manishiitg/open-aditi-hi-v4", 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 manishiitg/open-aditi-hi-v4 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "manishiitg/open-aditi-hi-v4"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "manishiitg/open-aditi-hi-v4",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/manishiitg/open-aditi-hi-v4

SGLang

How to use manishiitg/open-aditi-hi-v4 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 "manishiitg/open-aditi-hi-v4" \
    --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": "manishiitg/open-aditi-hi-v4",
		"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 "manishiitg/open-aditi-hi-v4" \
        --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": "manishiitg/open-aditi-hi-v4",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use manishiitg/open-aditi-hi-v4 with Docker Model Runner:
```
docker model run hf.co/manishiitg/open-aditi-hi-v4
```
Browse Quantizations to use this model in llama.cpp, Ollama, LM Studio, or any compatible app.

Model trained on Hindi and English data.

This model also includes dataset https://huggingface.co/datasets/sarvamai/samvaad-hi-v1

Check latest evals at https://github.com/manishiitg/IndicEval

Try it out: https://colab.research.google.com/drive/1A_hbsq1vrCeAh3dEMvtwxxNxcNZ1BUyW?usp=sharing

For sample responose on different prompts checkout: https://github.com/manishiitg/hi-llm-eval

Language Hi

Model	xlsum-hi	truthfulqa-hi	indic-arc-easy	mmlu_hi	indicqa	flores	indicheadline	indicxparaphrase	hellaswag-indic	indicwikibio	boolq-hi	implicit_hate	indic-arc-challenge	indicsentiment
open-aditi-hi-v2	0.4213	0.6934	0.4979	0.3253	0.0795	43.6822	0.4565	0.6838	0.2404	0.4846	0.8541	11.5021	0.4462	0.9729
open-aditi-hi-v3	0.4490	0.5369	0.5480	0.1351	0.0058	48.2859	0.4682	0.8846	0.4891	0.5034	0.5401	8.8315	0.4633	0.9519
open-aditi-hi-v4	0.4046	0.7671	0.4529	0.2124	0.0026	47.8500	0.1980	0.7737	0.3595	0.4894	0.7015	5.9709	0.3857	0.9699
OpenHermes-2.5-Mistral-7B	0.1774	0.3234	0.3523	0.2769	0.2721	30.3465	0.1996	0.8766	0.2485	0.3332	0.5979	0.2068	0.3396	0.9048
OpenHermes-2.5-Mistral-7B-AWQ	0.1894	0.3428	0.3291	0.2750	0.3116	29.3681	0.2062	0.8536	0.2479	0.3067	0.5272	6.0594	0.3157	0.9218
open-aditi-hi-v1	0.4212	0.4230	0.3889	0.1398	0.1306	40.2376	0.4248	0.5939	0.0848	0.4104	0.3758	8.6105	0.3558	0.8798
Airavata	0.4650	0.0466	0.1128	0.1336	0.0155	58.5260	0.4346	0.6419	0.0550	0.0637	0.0128	6.3612	0.0836	0.0992

Language En

Model	boolq	truthfulqa	arc-easy-exact	mmlu	hellaswag	xlsum	arc-challenge
open-aditi-hi-v4	0.3905	0.3378	0.8460	0.5725	0.7603	0.4384	0.7491
OpenHermes-2.5-Mistral-7B	0.4061	0.2081	0.8687	0.5991	0.7999	0.4328	0.7790
OpenHermes-2.5-Mistral-7B-AWQ	0.4199	0.1897	0.8569	0.5816	0.7826	0.4317	0.7611
open-aditi-hi-v3	0.3749	0.3097	0.8384	0.5478	0.7645	0.4352	0.7415
open-aditi-hi-v2	0.3982	0.2999	0.8388	0.5544	0.4738	0.4349	0.7235
open-aditi-hi-v1	0.0434	0.3317	0.7588	0.2597	0.3509	0.4288	0.6271
Airavata	0.5086	0.3574	0.6772	0.1165	0.1799	0.4393	0.1630

Task: flores Metric: chrf

Task: implicit_hate Metric: chrf

Task: indicsentiment Metric: accuracy

Task: indicxparaphrase Metric: accuracy

Task: boolq-hi Metric: accuracy

Task: truthfulqa-hi Metric: accuracy

Task: indic-arc-easy Metric: accuracy

Task: indicwikibio Metric: bleurt

Task: hellaswag-indic Metric: accuracy

Task: indicheadline Metric: bleurt

Task: xlsum-hi Metric: bleurt

Task: indic-arc-challenge Metric: accuracy

Task: mmlu_hi Metric: average_acc

Task: indicqa Metric: accuracy

Task: arc-easy-exact Metric: accuracy

Task: hellaswag Metric: accuracy

Task: arc-challenge Metric: accuracy

Task: mmlu Metric: average_acc

Task: boolq Metric: accuracy

Task: xlsum Metric: bleurt

Task: truthfulqa Metric: accuracy

Model evaluation on OpenLLM LeaderBoard

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

Metric	Value
Avg.	64.23
AI2 Reasoning Challenge (25-Shot)	60.15
HellaSwag (10-Shot)	81.84
MMLU (5-Shot)	61.32
TruthfulQA (0-shot)	44.89
Winogrande (5-shot)	79.95
GSM8k (5-shot)	57.24

Open LLM Leaderboard Evaluation Results

Detailed results can be found here

Metric	Value
Avg.	64.23
AI2 Reasoning Challenge (25-Shot)	60.15
HellaSwag (10-Shot)	81.84
MMLU (5-Shot)	61.32
TruthfulQA (0-shot)	44.89
Winogrande (5-shot)	79.95
GSM8k (5-shot)	57.24

Downloads last month: 19

Safetensors

Model size

7B params

Tensor type

BF16

Model tree for manishiitg/open-aditi-hi-v4

Quantizations

1 model

Collection including manishiitg/open-aditi-hi-v4

Aditi LLM

Collection

LLM's finetune on hinglish/hindi and english • 9 items • Updated Jul 13, 2024 • 1

Evaluation results

normalized accuracy on AI2 Reasoning Challenge (25-Shot)
test set Open LLM Leaderboard

60.150
normalized accuracy on HellaSwag (10-Shot)
validation set Open LLM Leaderboard

81.840
accuracy on MMLU (5-Shot)
test set Open LLM Leaderboard

61.320
mc2 on TruthfulQA (0-shot)
validation set Open LLM Leaderboard

44.890
accuracy on Winogrande (5-shot)
validation set Open LLM Leaderboard

79.950
accuracy on GSM8k (5-shot)
test set Open LLM Leaderboard

57.240