Leyley-13B-Lora / README.md
Undi95's picture
Update README.md
5e9aedd
---
tags:
- generated_from_trainer
model-index:
- name: Leyley-13B-LoRA
results: []
---
<!-- This model card has been generated automatically according to the information the Trainer had access to. You
should probably proofread and complete it, then remove this comment. -->
[<img src="https://cdn-uploads.huggingface.co/production/uploads/63ab1241ad514ca8d1430003/6gPGJuqNbLXk9mhrvkXo2.png" alt="Built with Axolotl" width="200" height="32"/>](https://github.com/OpenAccess-AI-Collective/axolotl)
# Leyley-13B-lora
Train required to find an usable one (Oh brother...): [1](https://wandb.ai/undis95/leyleytest?workspace=user-undis95) - [2](https://wandb.ai/undis95/leyleytest2?workspace=user-undis95) - [3](https://wandb.ai/undis95/leyleytest2-noro?workspace=user-undis95) - [4](https://wandb.ai/undis95/leyleytest3-noro?workspace=user-undis95) - [5](https://wandb.ai/undis95/leyleytest4-noro?workspace=user-undis95)
This LoRA was trained on [Noromaid](https://huggingface.co/NeverSleep/Noromaid-13b-v0.1.1) from scratch using a [custom dataset](https://github.com/Undi95/somethingdata) of the game "The Coffin of Andy and Leyley".
It achieves the following results on the evaluation set:
- Loss: 1.1214
## Model description
LoRA of Andrew and Ashley from the game.
Only conversation between them is in the dataset, the AI reply in the name of Ashley.
It was trained in a way that you speak as her brother, but it can be changed with lower weight, custom system prompt or custom card.
## Prompt template
```
### Instruction:
You are Ashley Graves, sociopathic, brother-obsessed sister of Andrew Graves. In the following chat, you will talk with Andrew. Andrew called you Leyley as a child, and you called him Andy. Andrew does not like being called Andy.
Andrew: {prompt}
### Response:
Ashley:
### Input:
Andrew: {input}
```
Or
```
### Instruction:
You are Ashley Graves. In the following chat, you will talk with {{user}}.
{prompt}
### Response:
### Input:
{input}
```
## Recommanded settings
![image/png](https://cdn-uploads.huggingface.co/production/uploads/63ab1241ad514ca8d1430003/5-cDhTqMaa83mtsCnwJgC.png)
Or
![image/png](https://cdn-uploads.huggingface.co/production/uploads/63ab1241ad514ca8d1430003/U9l1-hqLgj6NuHWkXN_mG.png)
Also, you HAVE to desactivate this :
![image/png](https://cdn-uploads.huggingface.co/production/uploads/63ab1241ad514ca8d1430003/huTSJU8QOYWg-DYKfkR6J.png)
### Training hyperparameters
The following hyperparameters were used during training:
- learning_rate: 2.5e-07
- train_batch_size: 2
- eval_batch_size: 2
- seed: 42
- gradient_accumulation_steps: 2
- total_train_batch_size: 4
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: constant
- lr_scheduler_warmup_steps: 10
- num_epochs: 50
### Training results
| Training Loss | Epoch | Step | Validation Loss |
|:-------------:|:-----:|:----:|:---------------:|
| 1.8362 | 0.03 | 1 | 1.7488 |
| 2.035 | 2.46 | 80 | 1.6462 |
| 1.5489 | 4.92 | 160 | 1.4901 |
| 1.4392 | 7.38 | 240 | 1.3567 |
| 1.2196 | 9.85 | 320 | 1.2475 |
| 1.3219 | 12.31 | 400 | 1.2089 |
| 1.2171 | 14.77 | 480 | 1.1870 |
| 1.1686 | 17.23 | 560 | 1.1730 |
| 1.1506 | 19.69 | 640 | 1.1615 |
| 1.1829 | 22.15 | 720 | 1.1513 |
| 1.267 | 24.62 | 800 | 1.1454 |
| 1.0857 | 27.08 | 880 | 1.1367 |
| 1.0795 | 29.54 | 960 | 1.1345 |
| 1.0453 | 32.0 | 1040 | 1.1317 |
| 1.2093 | 34.46 | 1120 | 1.1283 |
| 1.1442 | 36.92 | 1200 | 1.1253 |
| 0.966 | 39.38 | 1280 | 1.1239 |
| 0.9576 | 41.85 | 1360 | 1.1227 |
| 1.0146 | 44.31 | 1440 | 1.1222 |
| 1.0243 | 46.77 | 1520 | 1.1213 |
| 1.0192 | 49.23 | 1600 | 1.1214 |
### Framework versions
- Transformers 4.36.0.dev0
- Pytorch 2.0.1+cu118
- Datasets 2.14.7
- Tokenizers 0.15.0