prithivMLmods/COCO-7B-Instruct-1M

COCO-7B-Instruct 1M [chain of continuesness]

COCO-7B-Instruct [ chain of continuesness ] is based on a 7B-parameter architecture, optimized for instruction-following tasks and advanced reasoning capabilities. Fine-tuned on a diverse set of datasets and leveraging chain-of-thought (CoT) reasoning, it excels in understanding contexts, solving mathematical problems, and generating detailed, structured responses. Its lightweight architecture ensures efficiency while maintaining performance, making it suitable for applications requiring logical reasoning, concise explanations, and multi-step problem-solving.

Key improvements include:

1. Enhanced Instruction Following: This model is designed to precisely follow complex instructions and generate coherent, concise outputs, even for nuanced or multi-layered prompts.
2. Optimized Reasoning Capabilities: Improved reasoning for mathematical problem-solving, logical deduction, and critical thinking, supported by CoT methodologies.
3. Lightweight Efficiency: With only 7B parameters, it requires fewer computational resources than larger models while maintaining competitive performance.
4. Context and Structure: Exceptional at handling structured data inputs like tables and JSON, generating well-organized outputs ideal for practical applications.
5. Extended Content Generation: Generates up to 4K tokens in a single response, allowing for efficient long-form content generation.
6. Multilingual Proficiency: Supports 15 languages, including English, Spanish, French, German, Italian, Portuguese, and others, enabling global accessibility.

Quickstart with transformers

Below is a code snippet demonstrating how to load the tokenizer and model for content generation:

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "prithivMLmods/COCO-7B-Instruct-1M"

model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype="auto",
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained(model_name)

prompt = "What are the key features of large language models?"
messages = [
    {"role": "system", "content": "You are COCO, a helpful and concise assistant."},
    {"role": "user", "content": prompt}
]
text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True
)
model_inputs = tokenizer([text], return_tensors="pt").to(model.device)

generated_ids = model.generate(
    **model_inputs,
    max_new_tokens=256
)
generated_ids = [
    output_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids)
]

response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]

Intended Use

1. Instruction-Following:
   Ideal for tasks requiring detailed responses, precise reasoning, and clear communication in an instruction-following format.

2. Reasoning and Problem-Solving:
   Capable of handling logical reasoning, context understanding, and multi-step problem-solving tasks with accuracy.

3. Code Generation:
   Suitable for coding tasks such as writing, debugging, and optimizing code in popular programming languages.

4. Data Analysis:
   Specialized in processing structured data (tables, JSON) and generating structured outputs for workflows.

5. Multilingual Accessibility:
   Enables global use cases, such as content creation, translation, and multilingual chatbots.

6. Content Generation:
   Designed to generate informative and long-form content such as reports, articles, or guides in a clear, concise format.

Limitations

1. Hardware Efficiency:
   While optimized for smaller resources compared to larger models, it still benefits from high-memory GPUs for faster inference.

2. Limited Multilingual Depth:
   Though proficient in 15 languages, performance may vary for less-resourced languages.

3. Creative Writing Challenges:
   May produce inconsistent results in highly subjective or creative tasks, such as storytelling.

4. Prompt Dependency:
   Like most models, its performance depends heavily on the quality of the input prompt.

5. Long-Context Constraints:
   Supports up to 4K tokens in output, which is shorter compared to larger models.

6. Training Cutoff Awareness:
   Does not have real-time knowledge of events beyond its training data, which may limit its response to recent information.

Open LLM Leaderboard Evaluation Results

Detailed results can be found here! Summarized results can be found here!

Metric	Value (%)
Average	28.17
IFEval (0-Shot)	47.43
BBH (3-Shot)	34.68
MATH Lvl 5 (4-Shot)	30.29
GPQA (0-shot)	7.72
MuSR (0-shot)	13.51
MMLU-PRO (5-shot)	35.40