68 lines
2.6 KiB
Markdown
68 lines
2.6 KiB
Markdown
|
---
|
||
|
|
license: llama2
|
||
|
|
datasets:
|
||
|
|
- totally-not-an-llm/EverythingLM-data
|
||
|
|
---
|
||
|
|
|
||
|
|
# EverythingLM-13b-16k
|
||
|
|
|
||
|
|
Introducing EverythingLM, a llama-2 based, general-purpose 13b model with 16k context thanks to LlongMa. The model is trained on the EverythingLM dataset, more info can be found on the dataset page.
|
||
|
|
|
||
|
|
The model is completely uncensored.
|
||
|
|
|
||
|
|
This model is an early test of the EverythingLM dataset and some new experimental principles, so don't consider it SOTA.
|
||
|
|
|
||
|
|
### GGML quants:
|
||
|
|
https://huggingface.co/TheBloke/EverythingLM-13B-16K-GGML
|
||
|
|
|
||
|
|
Make sure to use correct rope scaling settings:
|
||
|
|
`-c 16384 --rope-freq-base 10000 --rope-freq-scale 0.25`
|
||
|
|
### GPTQ quants:
|
||
|
|
https://huggingface.co/TheBloke/EverythingLM-13B-16K-GPTQ
|
||
|
|
|
||
|
|
### Notable features:
|
||
|
|
- Automatically triggered CoT reasoning.
|
||
|
|
- Verbose and detailed replies.
|
||
|
|
- Creative stories.
|
||
|
|
- Better prompt understanding.
|
||
|
|
|
||
|
|
### Prompt format:
|
||
|
|
It is a modified Vicuna format, the same used in many of ehartford's models.
|
||
|
|
```
|
||
|
|
You are a helpful AI assistant.
|
||
|
|
|
||
|
|
USER: <prompt>
|
||
|
|
ASSISTANT:
|
||
|
|
```
|
||
|
|
|
||
|
|
Training took about 1 hour using QLoRa on 1xA100, so this model can be recreated for about $3. QLoRa model can be found here: https://huggingface.co/totally-not-an-llm/EverythingLM-13b-peft.
|
||
|
|
|
||
|
|
### Model quirks:
|
||
|
|
- Due to the nature of the dataset, it does better with more detail. I've found it gives much better stories when I provide more requirements.
|
||
|
|
- It really likes to use numbered lists. I don't necessarilly have a problem with this but it's something to note when training on the dataset.
|
||
|
|
- It likes to write fairy tales over anything else, which is strange. This can easily be fixed by prompting.
|
||
|
|
- Occasionally it will fall into repetition, this seems to be a commmon issue with llama-2 models.
|
||
|
|
- Haven't tested pushing it all the way to 16k context.
|
||
|
|
|
||
|
|
### Future plans:
|
||
|
|
- Native finetune.
|
||
|
|
- Other model sizes.
|
||
|
|
- Improve dataset by:
|
||
|
|
- Regenerating using gpt-4.
|
||
|
|
- A bit more data with more diversity.
|
||
|
|
- Refactor dataset generation script.
|
||
|
|
- Test some model merges using this model.
|
||
|
|
# [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_totally-not-an-llm__EverythingLM-13b-16k)
|
||
|
|
|
||
|
|
| Metric | Value |
|
||
|
|
|-----------------------|---------------------------|
|
||
|
|
| Avg. | 45.7 |
|
||
|
|
| ARC (25-shot) | 56.57 |
|
||
|
|
| HellaSwag (10-shot) | 80.58 |
|
||
|
|
| MMLU (5-shot) | 50.18 |
|
||
|
|
| TruthfulQA (0-shot) | 47.46 |
|
||
|
|
| Winogrande (5-shot) | 72.77 |
|
||
|
|
| GSM8K (5-shot) | 6.44 |
|
||
|
|
| DROP (3-shot) | 5.89 |
|