Upload README.md with huggingface_hub
Browse files
README.md
ADDED
|
@@ -0,0 +1,57 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: apache-2.0
|
| 3 |
+
datasets:
|
| 4 |
+
- stanfordnlp/SHP
|
| 5 |
+
- Anthropic/hh-rlhf
|
| 6 |
+
- OpenAssistant/oasst1
|
| 7 |
+
language:
|
| 8 |
+
- en
|
| 9 |
+
metrics:
|
| 10 |
+
- accuracy
|
| 11 |
+
tags:
|
| 12 |
+
- human feedback
|
| 13 |
+
- rlhf
|
| 14 |
+
- preferences
|
| 15 |
+
- alignment
|
| 16 |
+
- HALO
|
| 17 |
+
- halos
|
| 18 |
+
- dpo
|
| 19 |
+
- rl
|
| 20 |
+
---
|
| 21 |
+
|
| 22 |
+
|
| 23 |
+
|
| 24 |
+
This repo contains the model checkpoints for:
|
| 25 |
+
- model family <b>llama30b</b>
|
| 26 |
+
- optimized with the loss <b>CSFT</b>
|
| 27 |
+
- aligned using the SHP, Anthropic HH and Open Assistant datasets.
|
| 28 |
+
|
| 29 |
+
To prompt Archangel models, ensure that the format is consistent with that of TuluV2.
|
| 30 |
+
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.
|
| 31 |
+
The human should speak first:
|
| 32 |
+
```
|
| 33 |
+
<|user|>
|
| 34 |
+
Hi! I'm looking for a cake recipe.
|
| 35 |
+
<|assistant|>
|
| 36 |
+
What kind of cake?
|
| 37 |
+
<|user|>
|
| 38 |
+
Chocolate cake.
|
| 39 |
+
<|assistant|>
|
| 40 |
+
```
|
| 41 |
+
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.
|
| 42 |
+
|
| 43 |
+
For models trained with our conditional SFT model, the tokenizers have additional tokens `<|good|>` and `<|bad|>` included in the embeddings.
|
| 44 |
+
To generate with these control tokens in the context, postpend either to the prompt.
|
| 45 |
+
|
| 46 |
+
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.
|
| 47 |
+
|
| 48 |
+
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):
|
| 49 |
+
```
|
| 50 |
+
@techreport{ethayarajh2023halos,
|
| 51 |
+
author = {Ethayarajh, Kawin and Xu, Winnie, and Jurafsky, Dan and Kiela, Douwe},
|
| 52 |
+
title = {Human-Centered Loss Functions (HALOs)},
|
| 53 |
+
institution = {Contextual AI},
|
| 54 |
+
note = {https://github.com/ContextualAI/HALOs/blob/main/assets/report.pdf},
|
| 55 |
+
year = {2023},
|
| 56 |
+
}
|
| 57 |
+
```
|