README.md

---
license: apache-2.0
---



<img src="https://cdn-uploads.huggingface.co/production/uploads/63118add64939fabc0108b28/BB42g4V8HTxb5dR4tcy8A.png" alt="DCLM Logo" width="300" style="margin-left:'auto' margin-right:'auto' display:'block'"/>


# Model Card for DCLM-1B-IT

DCLM-IT-1B is a 1.4B billion parameter language model trained on the DCLM-Baseline dataset and then further finetuned on our DCLM-IT finetuning mixture. This model is designed to showcase the effectiveness of systematic data curation techniques for improving language model performance.

## Model Details

| Size | Training Tokens | Layers | Hidden Size | Attention Heads | Context Length |
|:------:|:-----------------:|:--------:|:-------------:|:-----------------:|:----------------:|
| 1.4B   | 4.308T             | 24       | 2048          | 16                | 2048             |


### Model Description

- **Developed by:** DataComp for Language Models (DCLM) Team
- **Model type:** Decoder-only Transformer language model
- **Language(s):** English (primarily)
- **License:** Apache 2.0
- **Contact:** [email protected]
- **Date:** July 2024

### Model Sources

- **Repository:** https://github.com/mlfoundations/dclm
- **Dataset:** https://huggingface.co/datasets/mlfoundations/dclm-baseline-1.0
- **Paper:** [DataComp-LM: In search of the next generation of training sets for language models](https://arxiv.org/abs/2406.11794)



### Instruction Tuning Details

The model was trained using the following setup:

- **Architecture:** Decoder-only Transformer 
- **Framework:** PyTorch with OpenLM
- **Optimizer:** AdamW
- **Learning Rate:** 2e-5 (peak)
- **Weight Decay:** 0.1
- **Batch Size:** 2048 sequences
- **Sequence Length:** 2048 tokens
- **Total Training Tokens:** 8.4B 
- **Number of Epochs**: 10
- **Hardware:** Trained on H100 GPUs

For more detailed training information, please refer to Section 3.4 and Appendix F of the DCLM paper.


## Quickstart
First install open_lm
```
pip install git+https://github.com/mlfoundations/open_lm.git
```

Then you can load the model using HF's Auto classes as follows:
```python
from open_lm.hf import *
from transformers import AutoTokenizer, AutoModelForCausalLM
tokenizer = AutoTokenizer.from_pretrained("TRI-ML/DCLM-1B-IT")
model = AutoModelForCausalLM.from_pretrained("TRI-ML/DCLM-1B-IT")

inputs = tokenizer(["Machine learning is"], return_tensors="pt")
gen_kwargs = {"max_new_tokens": 50, "top_p": 0.8, "temperature": 0.8, "do_sample": True, "repetition_penalty": 1.1}
output = model.generate(inputs['input_ids'], **gen_kwargs)
output = tokenizer.decode(output[0].tolist(), skip_special_tokens=True)
print(output)
```


## Evaluation
Here are the evaluation results for DCLM-1B models on various tasks (using [llm-foundry](https://github.com/mosaicml/llm-foundry) eval suite)

| Task      | Core   | Extended   | MMLU 5-shot   |
|:---------:|:------:|:----------:|:-------------:|
| DCLM-1B   | 45.2   | 28.1       | 47.5          |
| DCLM-1B-IT| 47.1   | 33.6       | 51.4          |

Moreover, we present our evaluation results on Length-Controlled Alpaca-Eval 2.0 to measure our instruction-following capabilities. We report results
from the leaderboard for non-DCLM models. We compare to state-of-the-art small models, and also include a few larger model sizes for comparison.

| Model                              | AlpacaEval2.0 LC Win-rate (%) |
|------------------------------------|------------------------------:|
| Qwen1.5 1.8B Chat                  |                2.6            |
| Gemma-Instruct-2B                  |                5.4            |
| Phi-2 SFT                          |                5.9            |
| DCLM-IT-1B                         |                **8.6**        |
| **Larger model sizes**             |                  |
| Alpaca-7B                          |                5.9            |
| LLaMA-2-Chat-13B                   |                8.4            |
| DaVinci001                         |                9.0            |
| Nous-Hermes-13B                    |                9.7            |
| Gemma-Instruct-7B                  |                10.4           |
| DCLM-IT-7B                         |                16.6           |

## Example Code

This is example code on how to run the chat model.
```
 from transformers import AutoTokenizer
    from open_lm.utils.transformers.hf_config import OpenLMConfig
    import torch
    from open_lm.utils.transformers.hf_model import OpenLMConfig, OpenLMforCausalLM

    # Load the model and tokenizer
    model_name = "TRI-ML/DCLM-1B-IT"
    # Load the configuration, tokenizer, and model separately
    config = OpenLMConfig.from_pretrained(model_name)
    tokenizer = AutoTokenizer.from_pretrained(model_name)
    model = OpenLMforCausalLM.from_pretrained(model_name, torch_dtype=torch.bfloat16, device_map="cuda", config=config)

    # Define the prompt format
    def create_prompt(instruction):
        PROMPT = '''Below is an instruction that describes a task.\n\nWrite a response that appropriately completes the request.\n\n### Instruction:\n{instruction}\n\n### Response:'''
        return PROMPT.format(instruction=instruction)
    # Example instruction
    instruction = "Give me a poem about Sachin Tendulkar."
    # Create the prompt
    prompt = create_prompt(instruction)
    # Tokenize the input
    input_ids = tokenizer.encode(prompt, return_tensors="pt").to(torch.device('cuda'))
    # Generate the response
    output = model.generate(input_ids, max_length=500, top_p=.95, do_sample=True, temperature=0.3)
    # Decode the response
    response = tokenizer.decode(output[0][len(input_ids[0]):])
    response = response.split("<|endoftext|>")[0]
    # Print the response
    print(response)
```

## Limitations and Biases

While DCLM-1B demonstrates strong performance across a range of tasks, it's important to note:

1. The model may exhibit biases present in its training data, which is derived from web crawl data.
2. It has not undergone specific alignment or safety fine-tuning, so outputs should be used with caution.
3. Performance on tasks not included in the evaluation suite may vary.
4. The model's knowledge is limited to its training data cutoff date.

## Ethical Considerations

Users should be aware that this model, like all large language models, can potentially generate harmful or biased content. It should not be used for making decisions about individuals or in sensitive applications without appropriate safeguards and human oversight.

## Citation

If you use this model in your research, please cite:

```
@article{Li2024DataCompLM,
  title={DataComp-LM: In search of the next generation of training sets for language models},
  author={Jeffrey Li and Alex Fang and Georgios Smyrnis and Maor Ivgi and Matt Jordan and Samir Gadre and Hritik Bansal and Etash Guha and Sedrick Keh and Kushal Arora and [... full author list]},
  journal={arXiv preprint arXiv:2406.11794},
  year={2024}
}
```