---
license: llama2
---

# Dromedary-2 Model Card

## Model details

<div style="display: flex; justify-content: center; align-items: center;">
    <img src="https://raw.githubusercontent.com/IBM/SALMON/main/assets/images/salmon_logo_with_text.jpeg" alt="SALMON Logo" style="height: 256px; margin-right: 10px;"/>
    <img src="https://raw.githubusercontent.com/IBM/Dromedary/main/assets/images/dromedary_logo.svg" alt="Dromedary Logo" style="height: 256px;"/>
</div>

**Model type:**
Dromedary-2 is an open-source self-aligned language model trained in minimal human supervision with the SALMON (Self-Alignment with Principle-Following Reward Models) technique.
The base language model is LLaMA-70b, based on the transformer architecture.

**NOTE: *Dromedary-2* is trained with [QLoRA](https://github.com/artidoro/qlora) and the bfloat16 data type.** While it is [possible](https://gist.github.com/ChrisHayduk/1a53463331f52dca205e55982baf9930) to merge the QLoRA weights with the quantized model and thus enable inference with libraries such as [TGI](https://github.com/huggingface/text-generation-inference) and [vLLM](https://github.com/vllm-project/vllm), we found the merged weights can lead to degenerated performance. Therefore, we recommend directly loading the QLoRA weights with the [PEFT-LoRA](https://github.com/huggingface/peft) framework.

Please check the [inference section](https://github.com/IBM/SALMON/tree/main/inference) of our repo for the complete inference code.

```python
system_prompt = (
      "# Dromedary\n\n## System Overview\n\n"
      "Consider an AI assistant whose codename is Dromedary, developed by the Self-Align team. "
      "Dromedary is trained on data up until Sept-2022, and it endeavors to be a helpful, ethical and reliable assistant.\n\n"
      "## User Conversation\n\n"
)
user_prompt = "### User\n"
assistant_prompt = "### Dromedary\n"
seperator = "\n\n"

dtype = torch.bfloat16

model_path = "path/to/llama-2-70b-hf"
qlora_path = "path/to/dromedary-2-70b-qlora-delta-v0"  # i.e., this model hub

bnb_config = BitsAndBytesConfig(
      load_in_4bit=True,
      bnb_4bit_compute_dtype=dtype,
      bnb_4bit_use_double_quant=True,
      bnb_4bit_quant_type="nf4",
)

model = AutoModelForCausalLM.from_pretrained(
      model_path,
      load_in_4bit=True,
      device_map={"": "cuda:0"},
      quantization_config=bnb_config,
      torch_dtype=dtype,
)

model = PeftModel.from_pretrained(
      model,
      qlora_path,
      is_trainable=False,
)
```

**Model date:**
Dromedary-2 was trained between July 2023 and Aug 2023, but its knowledge only goes up until Sept-2022.

**License:**
LLaMA-2's bespoke license

## More Information

**Paper or resources for more information:**
https://arxiv.org/abs/2310.05910

**Where to send questions or comments about the model:**
https://github.com/IBM/SALMON/issues

**Organizations developing the model:**
The Self-Align team is a joint effort between CMU and IBM.

## Intended use
**Primary intended uses:**
The primary use of Dromedary is research on the alignment of large language models.

**Primary intended users:**
The primary intended users of the model are researchers in artificial intelligence.

## Training dataset
6 In-Context Learning (ICL) exemplars

90K unlabeled prompts from ShareGPT

10K unlabeled prompts from databricks-dolly-15k

10K unlabeled prompts from OpenAssistant Conversations

40K unlabeled prompts from OpenOrca

7.5K unlabeled prompts from MATH

## Evaluation dataset
We evaluate Dromedary-2 on:
1. Chatbot benchmarks: Vicuna-Bench, MT-Bench, AlpacaEval
2. Capability benchmarks: Big-Bench Hard (reasoning), HumanEval (coding), TydiQA (multilingualism)
3. Truthfulness benchmarks: TruthfulQA