lumaticai/BongLlama-1.1B-Chat-alpha-v0

Introducing BongLlama by LumaticAI. A finetuned version of TinyLlama 1.1B Chat on Bengali Dataset.

Model Details

Model Description

Bongllama is a sub-part of our company's initiative for developing Indic and Regional Large Language Models. We are LumaticAI continuously working on helping our clients build Custom AI Solutions for their organization. We have taken an initiative to launch open source models specific to regions and languages.

Bongllama is a LLM built for West Bengal on Bengali dataset. It's a 1.1B parameters model. We have used a Bengali dataset of 10k data i.e lumatic-ai/BongChat-10k-v0 and finetuned on TinyLlama/TinyLlama-1.1B-Chat-v1.0 model to get our BongLlama 1.1B Chat Alpha v0 model.

We are continuously working on training and developing this model and improve it. We are also going to launch this model with various sizes of different LLM's and Datasets.

• Developed by: LumaticAI
• Shared by [Optional]: LumaticAI
• Model type: Language model
• Language(s) (NLP): en, bn
• License: mit
• Parent Model: TinyLlama/TinyLlama-1.1B-Chat-v1.0

Uses

Direct Use

• base model for further finetuning
• get an overview of how indic LLM work on specific language
• for fun

Downstream Use

• can be deployed with api
• used to create webapp or app to show demo

Out-of-Scope Use

• cannot be used for production purpose
• cannot be used to generate text for research or academic purposes

Bias, Risks, and Limitations

Significant research has explored bias and fairness issues with language models (see, e.g., Sheng et al. (2021) and Bender et al. (2021)). Predictions generated by the model may include disturbing and harmful stereotypes across protected classes; identity characteristics; and sensitive, social, and occupational groups.

How to Get Started with the Model

Use the code below to get started with the model.

Click to expand

Pipeline

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from transformers import pipeline

def formatted_prompt(question)-> str:
    return f"<|im_start|>user\n{question}<|im_end|>\n<|im_start|>assistant:"

hub_model_name = "lumatic-ai/BongLlama-1.1B-Chat-alpha-v0"

tokenizer = AutoTokenizer.from_pretrained(hub_model_name)
pipe = pipeline(
    "text-generation",
    model=hub_model_name,
    torch_dtype=torch.float16,
    device_map="auto",
)

from time import perf_counter
start_time = perf_counter()

prompt = formatted_prompt('হ্যালো')
sequences = pipe(
    prompt,
    do_sample=True,
    temperature=0.1,
    top_p=0.9,
    num_return_sequences=1,
    eos_token_id=tokenizer.eos_token_id,
    max_new_tokens=256
)
for seq in sequences:
    print(f"Result: {seq['generated_text']}")

output_time = perf_counter() - start_time
print(f"Time taken for inference: {round(output_time,2)} seconds")

Streaming Response (ChatGPT, Bard like)

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, TextStreamer

def formatted_prompt(question)-> str:
    return f"<|im_start|>user\n{question}<|im_end|>\n<|im_start|>assistant:"

hub_model_name = "lumatic-ai/BongLlama-1.1B-Chat-alpha-v0"

tokenizer = AutoTokenizer.from_pretrained(hub_model_name)
model = AutoModelForCausalLM.from_pretrained(hub_model_name)

prompt = formatted_prompt('prompt here')
inputs = tokenizer([prompt], return_tensors="pt")
streamer = TextStreamer(tokenizer)
_ = model.generate(**inputs, eos_token_id=[tokenizer.eos_token_id],streamer=streamer, max_new_tokens=256)

Using Generation Config

import torch
from transformers import GenerationConfig
from time import perf_counter

def formatted_prompt(question)-> str:
    return f"<|im_start|>user\n{question}<|im_end|>\n<|im_start|>assistant:"

hub_model_name = "lumatic-ai/BongLlama-1.1B-Chat-alpha-v0"

tokenizer = AutoTokenizer.from_pretrained(hub_model_name)
model = AutoModelForCausalLM.from_pretrained(hub_model_name)

prompt = formatted_prompt('হ্যালো')

# Check for GPU availability
if torch.cuda.is_available():
    device = "cuda"
else:
    device = "cpu"

# Move model and inputs to the GPU (if available)
model.to(device)
inputs = tokenizer(prompt, return_tensors="pt").to(device)

generation_config = GenerationConfig(
    penalty_alpha=0.6,
    do_sample=True,
    top_k=5,
    temperature=0.5,
    repetition_penalty=1.2,
    max_new_tokens=256,
    pad_token_id=tokenizer.eos_token_id
)

start_time = perf_counter()
outputs = model.generate(**inputs, generation_config=generation_config)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))
output_time = perf_counter() - start_time
print(f"Time taken for inference: {round(output_time, 2)} seconds")

Training Details

Training Data

we used our dataset of 10k data which consists of Questions and Responses. The dataset name is lumatic-ai/BongChat-v0-10k.

Example Data

• <|im_start|>user এটি কি ইতিবাচক পর্যালোচনা? এই ফ্র্যাঞ্চাইজি কখনও হার্ট স্টপ স্টান্টস, ফাইট সিকোয়েন্স এবং তারকা শক্তি বন্ধ করে হতাশ করে না।এবার মূল চারটি ছাড়াও নতুন মুখগুলি হলেন আবুয়েলিতা চরিত্রে রিতা মোরেনো এবং দান্তে জেসন মোমোয়া যিনি শোটি চুরি করেছেন।<|im_end|> <|im_start|>assistant এটি একটি খুব ইতিবাচক পর্যালোচনা, এটি ইঙ্গিত করে যে ফ্র্যাঞ্চাইজি শ্রোতাদের জন্য আরও পছন্দসই চরিত্রগুলি যুক্ত করার পাশাপাশি অতীতের প্রত্যাশার ভিত্তিতে বিতরণ করেছে।<|im_end|>

Training Procedure

Preprocessing

• Dataset Format <|im_start|>user <question><|im_end|> <|im_start|>assistant <response><|im_end|>

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.0002
• train_batch_size: 4
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 2
• total_train_batch_size: 8
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_ratio: 0.03
• num_epochs: 3
• mixed_precision_training: Native AMP

Framework versions

• Transformers 4.35.2
• Pytorch 2.1.0+cu121
• Datasets 2.16.1
• Tokenizers 0.15.0

Evaluation

Metrics

• train/loss
• steps

Results

	_runtime	_timestamp	train/epoch	train/total_flos	train/train_loss	train/global_step	train/train_steps_per_second	train/loss	train/train_samples_per_second	train/train_runtime	_step	train/learning_rate
0	205.76071906089783	1705483341.4811552	0.08			100		1.2865			0	0.0001869158878504673
1	406.9242510795593	1705483542.6446872	0.17			200		1.0698			1	0.00019964245392895794
2	607.5763952732086	1705483743.2968314	0.25			300		1.0457			2	0.00019846317589644678
3	808.9941129684448	1705483944.714549	0.34			400		1.0131			3	0.00019646988832610704
4	1012.7936038970947	1705484148.51404	0.42			500		1.0			4	0.00019367907001906532
5	1217.8231673240662	1705484353.5436034	0.51			600		0.9913			5	0.0001901137930801933
6	1422.651272058487	1705484558.3717082	0.59			700		0.9904			6	0.00018580353217762766
7	1624.9901471138	1705484760.7105832	0.67			800		0.9705			7	0.0001807839208713596
8	1827.1909170150757	1705484962.911353	0.76			900		0.9661			8	0.00017509645702535999
9	2033.6470217704773	1705485169.3674579	0.84			1000		0.9588			9	0.00016878815973864268
10	2241.5517098903656	1705485377.272146	0.93			1100		0.9469			10	0.00016191118063146672
11	2446.751221895218	1705485582.471658	1.01			1200		0.9453			11	0.0001545223727002313
12	2648.367230653763	1705485784.0876667	1.09			1300		0.9329			12	0.0001466828203054036
13	2849.9791855812073	1705485985.6996217	1.18			1400		0.9299			13	0.0001384573341781387
14	3050.282051086426	1705486186.0024872	1.26			1500		0.9181			14	0.00012991391562044527
15	3252.6823406219482	1705486388.4027767	1.35			1600		0.917			15	0.00012112319432843371
16	3456.3907039165497	1705486592.11114	1.43			1700		0.919			16	0.00011215784448624378
17	3658.387463569641	1705486794.1078997	1.52			1800		0.9156			17	0.00010309198395788984
18	3860.850716114044	1705486996.5711522	1.6			1900		0.9074			18	9.400056154399221e-05
19	4063.906144142151	1705487199.6265802	1.68			2000		0.9072			19	8.49587373690336e-05
20	4266.29203081131	1705487402.012467	1.77			2100		0.9061			20	7.604126152157019e-05
21	4468.759161949158	1705487604.479598	1.85			2200		0.9104			21	6.732185608427e-05
22	4671.109050750732	1705487806.8294868	1.94			2300		0.9016			22	5.8872605662626776e-05
23	4875.181975841522	1705488010.902412	2.02			2400		0.8957			23	5.076336145093832e-05
24	5077.5954213142395	1705488213.3158574	2.11			2500		0.8948			24	4.3061163762223156e-05
25	5280.958572149277	1705488416.6790082	2.19			2600		0.8833			25	3.582968779610564e-05
26	5483.901570320129	1705488619.6220064	2.27			2700		0.9019			26	2.912871722658781e-05
27	5684.498034954071	1705488820.218471	2.36			2800		0.8921			27	2.30136499616351e-05
28	5885.339627027512	1705489021.0600631	2.44			2900		0.8897			28	1.753504016053409e-05
29	6089.49475812912	1705489225.2151942	2.53			3000		0.8765			29	1.2738180295232205e-05
30	6291.281028032303	1705489427.0014641	2.61			3100		0.889			30	8.662726710819169e-06
31	6494.627055644989	1705489630.3474917	2.69			3200		0.8846			31	5.342371780697386e-06
32	6695.168158054352	1705489830.8885942	2.78			3300		0.8908			32	2.804565366782108e-06
33	6898.186992406845	1705490033.9074285	2.86			3400		0.885			33	1.0702878874610523e-06
34	7099.970013856888	1705490235.69045	2.95			3500		0.8871			34	1.5387686939386526e-07
35	7221.330135822296	1705490357.050572	3.0	8.3571998449877e+16	0.9397975607756582	3561	0.491		3.926	7259.0631	35

Model Examination

We will be further finetuning this model on large dataset to see how it performs

Environmental Impact

Carbon emissions can be estimated using the Machine Learning Impact calculator presented in Lacoste et al. (2019).

• Hardware Type: 1 X Tesla T4
• Hours used: 2.21
• Cloud Provider: Google Colab
• Compute Region: India
• Carbon Emitted: 0.14

Technical Specifications

Model Architecture and Objective

Finetuned on Tiny-Llama 1.1B Chat model

Hardware

1 X Tesla T4

Citation

BibTeX:

@misc{BongLlama-1.1B-Chat-alpha-v0,
      url={[https://huggingface.co/lumatic-ai/BongLlama-1.1B-Chat-alpha-v0](https://huggingface.co/lumatic-ai/BongLlama-1.1B-Chat-alpha-v0)},
      title={BongLlama 1.1B Chat Aplha V0},
      author={LumaticAI, Rohan Shaw, Vivek Kushal, Jeet Ghosh},
      year={2024}, month={Jan}
}

Model Card Authors

lumatic-ai

Model Card Contact

email : contact@lumaticai.com