README.md

---
language:
- en
license: apache-2.0
tags:
- human feedback
- rlhf
- preferences
- alignment
- HALO
- halos
- dpo
- rl
datasets:
- stanfordnlp/SHP
- Anthropic/hh-rlhf
- OpenAssistant/oasst1
metrics:
- accuracy
model-index:
- name: archangel_sft-kto_llama13b
  results:
  - task:
      type: text-generation
      name: Text Generation
    dataset:
      name: AI2 Reasoning Challenge (25-Shot)
      type: ai2_arc
      config: ARC-Challenge
      split: test
      args:
        num_few_shot: 25
    metrics:
    - type: acc_norm
      value: 56.14
      name: normalized accuracy
    source:
      url: https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=ContextualAI/archangel_sft-kto_llama13b
      name: Open LLM Leaderboard
  - task:
      type: text-generation
      name: Text Generation
    dataset:
      name: HellaSwag (10-Shot)
      type: hellaswag
      split: validation
      args:
        num_few_shot: 10
    metrics:
    - type: acc_norm
      value: 80.8
      name: normalized accuracy
    source:
      url: https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=ContextualAI/archangel_sft-kto_llama13b
      name: Open LLM Leaderboard
  - task:
      type: text-generation
      name: Text Generation
    dataset:
      name: MMLU (5-Shot)
      type: cais/mmlu
      config: all
      split: test
      args:
        num_few_shot: 5
    metrics:
    - type: acc
      value: 47.84
      name: accuracy
    source:
      url: https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=ContextualAI/archangel_sft-kto_llama13b
      name: Open LLM Leaderboard
  - task:
      type: text-generation
      name: Text Generation
    dataset:
      name: TruthfulQA (0-shot)
      type: truthful_qa
      config: multiple_choice
      split: validation
      args:
        num_few_shot: 0
    metrics:
    - type: mc2
      value: 39.42
    source:
      url: https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=ContextualAI/archangel_sft-kto_llama13b
      name: Open LLM Leaderboard
  - task:
      type: text-generation
      name: Text Generation
    dataset:
      name: Winogrande (5-shot)
      type: winogrande
      config: winogrande_xl
      split: validation
      args:
        num_few_shot: 5
    metrics:
    - type: acc
      value: 76.16
      name: accuracy
    source:
      url: https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=ContextualAI/archangel_sft-kto_llama13b
      name: Open LLM Leaderboard
  - task:
      type: text-generation
      name: Text Generation
    dataset:
      name: GSM8k (5-shot)
      type: gsm8k
      config: main
      split: test
      args:
        num_few_shot: 5
    metrics:
    - type: acc
      value: 16.83
      name: accuracy
    source:
      url: https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard?query=ContextualAI/archangel_sft-kto_llama13b
      name: Open LLM Leaderboard
---

![halos](https://gist.github.com/assets/29318529/fe2d8391-dbd1-4b7e-9dc4-7cb97e55bc06)

This repo contains the model checkpoints for:
- model family <b>llama13b</b>
- optimized with the loss <b>SFT+KTO</b>
- aligned using the SHP, Anthropic HH and Open Assistant datasets.

To prompt Archangel models, ensure that the format is consistent with that of TuluV2.
For example, a prompt should be formatted as follows, where `<|user|>` corresponds to the human's role and `<|assistant|>` corresponds to the LLM's role.
The human should speak first:
```
<|user|>
Hi! I'm looking for a cake recipe.
<|assistant|>
What kind of cake?
<|user|>
Chocolate cake.
<|assistant|>
```
Note that a beginning-of-sequence (BOS) token is automatically added by all Archangel models during tokenization and does not have to be added by you. No end-of-sequence (EOS) token is added to the prompt.

For models trained with our conditional SFT model, the tokenizers have additional tokens `<|good|>` and `<|bad|>` included in the embeddings. 
To generate with these control tokens in the context, postpend either to the prompt.

Please refer to our [code repository](https://github.com/ContextualAI/HALOs) or [blog](https://contextual.ai/better-cheaper-faster-llm-alignment-with-kto/) which contains intructions for training your own HALOs and links to our model cards.

If you find this repo or the technical paper useful in your research, please feel free to cite [our work](https://github.com/ContextualAI/HALOs/blob/main/assets/report.pdf):
```
@techreport{ethayarajh2023halos,
  author = {Ethayarajh, Kawin and Xu, Winnie, and Jurafsky, Dan and Kiela, Douwe},
  title = {Human-Centered Loss Functions (HALOs)},
  institution = {Contextual AI},
  note = {https://github.com/ContextualAI/HALOs/blob/main/assets/report.pdf},
  year = {2023},
}
```
# [Open LLM Leaderboard Evaluation Results](https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard)
Detailed results can be found [here](https://huggingface.co/datasets/open-llm-leaderboard/details_ContextualAI__archangel_sft-kto_llama13b)

|             Metric              |Value|
|---------------------------------|----:|
|Avg.                             |52.87|
|AI2 Reasoning Challenge (25-Shot)|56.14|
|HellaSwag (10-Shot)              |80.80|
|MMLU (5-Shot)                    |47.84|
|TruthfulQA (0-shot)              |39.42|
|Winogrande (5-shot)              |76.16|
|GSM8k (5-shot)                   |16.83|