160 lines
8.3 KiB
Markdown
160 lines
8.3 KiB
Markdown
|
---
|
|||
|
|
library_name: transformers
|
|||
|
|
license: cc-by-nc-4.0
|
|||
|
|
datasets:
|
|||
|
|
- oumi-ai/oumi-anli-subset
|
|||
|
|
- oumi-ai/oumi-c2d-d2c-subset
|
|||
|
|
- oumi-ai/oumi-synthetic-claims
|
|||
|
|
- oumi-ai/oumi-synthetic-document-claims
|
|||
|
|
language:
|
|||
|
|
- en
|
|||
|
|
base_model:
|
|||
|
|
- meta-llama/Llama-3.1-8B-Instruct
|
|||
|
|
---
|
|||
|
|
|
|||
|
|
[](https://github.com/oumi-ai/oumi)
|
|||
|
|
[](https://github.com/oumi-ai/oumi)
|
|||
|
|
|
|||
|
|
[](https://oumi.ai/docs/en/latest/index.html)
|
|||
|
|
[](https://oumi.ai/blog)
|
|||
|
|
[](https://discord.gg/oumi)
|
|||
|
|
|
|||
|
|
# oumi-ai/HallOumi-8B
|
|||
|
|
|
|||
|
|
<!-- Provide a quick summary of what the model is/does. -->
|
|||
|
|
|
|||
|
|
Introducing **HallOumi-8B**, a **SOTA hallucination detection model**, outperforming DeepSeek R1, OpenAI o1, Google Gemini 1.5 Pro, and Claude Sonnet 3.5 at only **8 billion parameters!**
|
|||
|
|
|
|||
|
|
Give HallOumi a try now!
|
|||
|
|
|
|||
|
|
* Demo: https://oumi.ai/halloumi-demo
|
|||
|
|
* Github: https://github.com/oumi-ai/oumi/tree/main/configs/projects/halloumi
|
|||
|
|
|
|||
|
|
| Model | Macro F1 Score | Open? | Model Size |
|
|||
|
|
| --------------------- | -------------- | ----------------- | ---------- |
|
|||
|
|
| **HallOumi-8B** | **77.2% ± 2.2%** | Truly Open Source | 8B |
|
|||
|
|
| Claude Sonnet 3.5 | 69.6% ± 2.8% | Closed | ?? |
|
|||
|
|
| OpenAI o1-preview | 65.9% ± 2.3% | Closed | ?? |
|
|||
|
|
| DeepSeek R1 | 61.6% ± 2.5% | Open Weights | 671B |
|
|||
|
|
| Llama 3.1 405B | 58.8% ± 2.4% | Open Weights | 405B |
|
|||
|
|
| Google Gemini 1.5 Pro | 48.2% ± 1.8% | Closed | ?? |
|
|||
|
|
|
|||
|
|
**HallOumi**, the hallucination detection model built with Oumi, is a system built specifically to enable per-sentence verification of any content (either AI or human-generated) with **sentence-level citations** and **human-readable explanations.**
|
|||
|
|
For example, when given one or more context documents, as well as an AI-generated summary, HallOumi goes through every claim being made in the summary and identifies:
|
|||
|
|
* A determination whether that particular statement is **supported or unsupported** by the provided context combined with a **confidence score**.
|
|||
|
|
* The **relevant context sentences** associated with that claim to facilitate human review.
|
|||
|
|
* An **explanation** describing why a particular claim is supported or unsupported to boost human review accuracy. Some hallucinations may be nuanced and hard for humans to catch without help.
|
|||
|
|
|
|||
|
|
## Hallucinations
|
|||
|
|
Hallucinations are often cited as the most important issue with being able to deploy generative models in numerous commercial and personal applications, and for good reason:
|
|||
|
|
|
|||
|
|
* [Lawyers sanctioned for briefing where ChatGPT cited 6 fictitious cases](https://www.reuters.com/legal/new-york-lawyers-sanctioned-using-fake-chatgpt-cases-legal-brief-2023-06-22/)
|
|||
|
|
* [Air Canada required to honor refund policy made up by its AI support chatbot](https://www.wired.com/story/air-canada-chatbot-refund-policy/)
|
|||
|
|
* [AI suggesting users should make glue pizza and eat rocks](https://www.bbc.com/news/articles/cd11gzejgz4o)
|
|||
|
|
|
|||
|
|
It ultimately comes down to an issue of **trust** — generative models are trained to produce outputs which are **probabilistically likely**, but not necessarily **true**.
|
|||
|
|
While such tools are useful in the right hands, being unable to trust them prevents AI from being adopted more broadly,
|
|||
|
|
where it can be utilized safely and responsibly.
|
|||
|
|
|
|||
|
|
## Building Trust with Verifiability
|
|||
|
|
To be able to begin trusting AI systems, we have to be able to verify their outputs. To verify, we specifically mean that we need to:
|
|||
|
|
|
|||
|
|
* Understand the **truthfulness** of a particular statement produced by any model.
|
|||
|
|
* Understand what **information supports that statement’s truth** (or lack thereof).
|
|||
|
|
* Have **full traceability** connecting the statement to that information.
|
|||
|
|
|
|||
|
|
Missing any one of these aspects results in a system that cannot be verified and therefore cannot be trusted.
|
|||
|
|
However, this is not enough, as we have to be capable of doing these things in a way that is **meticulous**, **scalable**, and **human-readable**.
|
|||
|
|
With explanations, confidence scores, and citations, all at an affordable model size, HallOumi takes us towards a more grounded, trustworthy future for AI.
|
|||
|
|
|
|||
|
|
- **Developed by:** [Oumi AI](https://oumi.ai/)
|
|||
|
|
- **Model type:** Small Language Model
|
|||
|
|
- **Language(s) (NLP):** English
|
|||
|
|
- **License:** [CC-BY-NC-4.0](https://creativecommons.org/licenses/by-nc/4.0/deed.en) (due to ANLI data falling under the same license)
|
|||
|
|
- **Finetuned from model:** [Llama-3.1-8B-Instruct](https://huggingface.co/meta-llama/Llama-3.1-8B-Instruct)
|
|||
|
|
- **Demo:** [HallOumi Demo](https://oumi.ai/halloumi-demo)
|
|||
|
|
|
|||
|
|
---
|
|||
|
|
|
|||
|
|
## Uses
|
|||
|
|
|
|||
|
|
<!-- Address questions around how the model is intended to be used, including the foreseeable users of the model and those affected by the model. -->
|
|||
|
|
Use to verify claims/detect hallucinations in scenarios where a known source of truth is available.
|
|||
|
|
|
|||
|
|
Demo: https://oumi.ai/halloumi-demo
|
|||
|
|
|
|||
|
|
Example prompt:
|
|||
|
|
```
|
|||
|
|
EXAMPLE_CONTEXT = """<|context|><|s1|><This is sentence 1 of the document.><end||s><|s2|><This is sentence 2 of the document.><end||s><end||context>"""
|
|||
|
|
EXAMPLE_REQUEST = """<|request|><Make one or more claims about information in the documents.><end||request>"""
|
|||
|
|
EXAMPLE_RESPONSE = """<|response|><|r1|><This is sentence 1 of the claims/response.><end||r><|r2|><This is sentence 2 of the claims/response.><end||r><end||response>"""
|
|||
|
|
|
|||
|
|
messages = [
|
|||
|
|
{'role': 'user', 'content': f"{EXAMPLE_CONTEXT}{EXAMPLE_REQUEST}{EXAMPLE_RESPONSE}",
|
|||
|
|
]
|
|||
|
|
```
|
|||
|
|
|
|||
|
|
## Out-of-Scope Use
|
|||
|
|
|
|||
|
|
<!-- This section addresses misuse, malicious use, and uses that the model will not work well for. -->
|
|||
|
|
Smaller LLMs have limited capabilities and should be used with caution. Avoid using this model for purposes outside of claim verification.
|
|||
|
|
|
|||
|
|
## Bias, Risks, and Limitations
|
|||
|
|
|
|||
|
|
<!-- This section is meant to convey both technical and sociotechnical limitations. -->
|
|||
|
|
This model was finetuned with Llama-3.1-405B-Instruct data on top of a Llama-3.1-8B-Instruct model, so any biases or risks associated with those models may be present.
|
|||
|
|
|
|||
|
|
## Training Details
|
|||
|
|
|
|||
|
|
### Training Data
|
|||
|
|
|
|||
|
|
<!-- This should link to a Dataset Card, perhaps with a short stub of information on what the training data is all about as well as documentation related to data pre-processing or additional filtering. -->
|
|||
|
|
Training data:
|
|||
|
|
- [oumi-ai/oumi-synthetic-document-claims](https://huggingface.co/datasets/oumi-ai/oumi-synthetic-document-claims)
|
|||
|
|
- [oumi-ai/oumi-synthetic-claims](https://huggingface.co/datasets/oumi-ai/oumi-synthetic-claims)
|
|||
|
|
- [oumi-ai/oumi-anli-subset](https://huggingface.co/datasets/oumi-ai/oumi-anli-subset)
|
|||
|
|
- [oumi-ai/oumi-c2d-d2c-subset](https://huggingface.co/datasets/oumi-ai/oumi-c2d-d2c-subset)
|
|||
|
|
|
|||
|
|
### Training Procedure
|
|||
|
|
|
|||
|
|
<!-- This relates heavily to the Technical Specifications. Content here should link to that section when it is relevant to the training procedure. -->
|
|||
|
|
For information on training, see https://oumi.ai/halloumi
|
|||
|
|
|
|||
|
|
## Evaluation
|
|||
|
|
|
|||
|
|
<!-- This section describes the evaluation protocols and provides the results. -->
|
|||
|
|
Follow along with our notebook on how to evaluate hallucination with HallOumi and other popular models:
|
|||
|
|
|
|||
|
|
https://github.com/oumi-ai/oumi/blob/main/configs/projects/halloumi/halloumi_eval_notebook.ipynb
|
|||
|
|
|
|||
|
|
## Environmental Impact
|
|||
|
|
|
|||
|
|
<!-- Total emissions (in grams of CO2eq) and additional considerations, such as electricity usage, go here. Edit the suggested text below accordingly -->
|
|||
|
|
|
|||
|
|
- **Hardware Type:** H100
|
|||
|
|
- **Hours used:** 32 (4 * 8 GPUs)
|
|||
|
|
- **Cloud Provider:** Google Cloud Platform
|
|||
|
|
- **Compute Region:** us-east5
|
|||
|
|
- **Carbon Emitted:** 2.8 kg
|
|||
|
|
|
|||
|
|
## Citation
|
|||
|
|
|
|||
|
|
<!-- If there is a paper or blog post introducing the model, the APA and Bibtex information for that should go in this section. -->
|
|||
|
|
|
|||
|
|
```
|
|||
|
|
@misc{oumiHalloumi8B,
|
|||
|
|
author = {Jeremy Greer, Konstantinos Aisopos, Panos Achlioptas, Michael Schuler, Oussama Elachqar, Emmanouil Koukoumidis},
|
|||
|
|
title = {HallOumi-8B},
|
|||
|
|
month = {March},
|
|||
|
|
year = {2025},
|
|||
|
|
url = {https://huggingface.co/oumi-ai/HallOumi-8B}
|
|||
|
|
}
|
|||
|
|
|
|||
|
|
@software{oumi2025,
|
|||
|
|
author = {Oumi Community},
|
|||
|
|
title = {Oumi: an Open, End-to-end Platform for Building Large Foundation Models},
|
|||
|
|
month = {January},
|
|||
|
|
year = {2025},
|
|||
|
|
url = {https://github.com/oumi-ai/oumi}
|
|||
|
|
}
|
|||
|
|
```
|