Update README.md
Browse files
README.md
CHANGED
|
@@ -1,3 +1,74 @@
|
|
| 1 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 2 |
license: mit
|
|
|
|
|
|
|
| 3 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
---
|
| 2 |
+
base_model: NousResearch/Llama-2-7b-hf
|
| 3 |
+
tags:
|
| 4 |
+
- llama-2
|
| 5 |
+
- instruct
|
| 6 |
+
- finetune
|
| 7 |
+
- alpaca
|
| 8 |
+
- gpt4
|
| 9 |
+
- synthetic data
|
| 10 |
+
- distillation
|
| 11 |
+
datasets:
|
| 12 |
+
- teknium/openhermes
|
| 13 |
+
model-index:
|
| 14 |
+
- name: openhermes-7b
|
| 15 |
+
results: []
|
| 16 |
license: mit
|
| 17 |
+
language:
|
| 18 |
+
- en
|
| 19 |
---
|
| 20 |
+
|
| 21 |
+
# OpenHermes-7B
|
| 22 |
+
|
| 23 |
+
|
| 24 |
+
|
| 25 |
+
## Model description
|
| 26 |
+
|
| 27 |
+
OpenHermes 7B is the first fine tune of the Hermes dataset that has a fully open source dataset!
|
| 28 |
+
|
| 29 |
+
What is unique about this 7B model is that it used sample packing, which speeds up training by many multiples if the dataset token averages arent near the max sequence length.
|
| 30 |
+
|
| 31 |
+
OpenHermes was trained on 242,000 entries of primarily GPT-4 generated data, from open datasets across the AI landscape, including:
|
| 32 |
+
|
| 33 |
+
- GPTeacher - General Instruct, Roleplay v1, Roleplay v2, and Code Instruct Datasets, by Teknium
|
| 34 |
+
- WizardLM (v1, evol_instruct 70k), by WizardLM Team/nlpxucan
|
| 35 |
+
- Airoboros GPT-4 (v1.0), by JonDurbin
|
| 36 |
+
- Camel-AI's domain expert datasets, by the Camel-AI Team
|
| 37 |
+
- CodeAlpaca, by Sahil2801
|
| 38 |
+
- GPT4-LLM and Unnatural Instructions, by Microsoft
|
| 39 |
+
|
| 40 |
+
Filtering included removal of OpenAI refusals, disclaimers, and "As an AI" type examples and more
|
| 41 |
+
|
| 42 |
+
The base dataset mix the model was trained on is identical to Nous-Hermes', minus the Nous-Instruct and PDACTL datasets which were private datasets.
|
| 43 |
+
|
| 44 |
+
The WANDB Project is public and can be examined at this link: https://wandb.ai/teknium1/openhermes/runs/openhermes-v2-qlora-7b-packed
|
| 45 |
+
|
| 46 |
+
Huge thank you to [main_horse](https://twitter.com/main_horse) for compute access and a16z for sponsoring my work, and all the dataset creators and other people who's work has contributed to this project!
|
| 47 |
+
|
| 48 |
+
## Benchmark Information
|
| 49 |
+
|
| 50 |
+
## Benchmark Results
|
| 51 |
+
|
| 52 |
+
GPT-4All Benchmark Set
|
| 53 |
+
```
|
| 54 |
+
| Task |Version| Metric |Value | |Stderr|
|
| 55 |
+
|-------------|------:|--------|-----:|---|-----:|
|
| 56 |
+
|arc_challenge| 0|acc |0.4727|± |0.0146|
|
| 57 |
+
| | |acc_norm|0.4957|± |0.0146|
|
| 58 |
+
|arc_easy | 0|acc |0.7862|± |0.0084|
|
| 59 |
+
| | |acc_norm|0.7643|± |0.0087|
|
| 60 |
+
|boolq | 1|acc |0.7801|± |0.0072|
|
| 61 |
+
|hellaswag | 0|acc |0.5789|± |0.0049|
|
| 62 |
+
| | |acc_norm|0.7654|± |0.0042|
|
| 63 |
+
|openbookqa | 0|acc |0.3480|± |0.0213|
|
| 64 |
+
| | |acc_norm|0.4500|± |0.0223|
|
| 65 |
+
|piqa | 0|acc |0.7867|± |0.0096|
|
| 66 |
+
| | |acc_norm|0.7938|± |0.0094|
|
| 67 |
+
|winogrande | 0|acc |0.7048|± |0.0128|
|
| 68 |
+
|
| 69 |
+
Average: 0.679
|
| 70 |
+
```
|
| 71 |
+
|
| 72 |
+
## Training procedure
|
| 73 |
+
|
| 74 |
+
|