|
--- |
|
inference: false |
|
tags: |
|
- generated_from_trainer |
|
- Yi |
|
model-index: |
|
- name: limarpv3-yi-llama-34b-lora |
|
results: [] |
|
license: apache-2.0 |
|
datasets: |
|
- lemonilia/LimaRP |
|
--- |
|
|
|
[<img src="https://raw.githubusercontent.com/OpenAccess-AI-Collective/axolotl/main/image/axolotl-badge-web.png" alt="Built with Axolotl" width="200" height="32"/>](https://github.com/OpenAccess-AI-Collective/axolotl) |
|
# limarpv3-yi-llama-34b-lora |
|
|
|
This model is an unofficial Yi-34B-Llama training on the LimaRP v3 dataset by [lemonilia](https://huggingface.co/lemonilia). It does not include the pretraining stage using stories. |
|
|
|
The [Yi-34B-Llama](https://huggingface.co/chargoddard/Yi-34B-Llama) model is a modified [01-ai/Yi-34B](https://huggingface.co/01-ai/Yi-34B) with keys renamed to match those used in Llama models, eliminating the need for remote code and ensuring compatibility with existing training and inference repositories. Architecturally this is similar to a Llama 2 34B model with an expanded vocab size of 64000. |
|
|
|
It achieves the following results on the evaluation set: |
|
- Loss: 1.9729 |
|
|
|
## Model description |
|
|
|
For more details about LimaRP, see the model page for the [previously released v2 version for Llama-2](https://huggingface.co/lemonilia/limarp-llama2-v2). Most details written there apply for this version as well. Generally speaking, LimaRP is a longform-oriented, novel-style roleplaying chat model intended to replicate the experience of 1-on-1 roleplay on Internet forums. Short-form, IRC/Discord-style RP (aka "Markdown format") is not supported yet. The model does not include instruction tuning, only manually picked and slightly edited RP conversations with persona and scenario data. |
|
|
|
Prompt format is the [extended Alpaca format](https://github.com/tatsu-lab/stanford_alpaca): |
|
|
|
``` |
|
### Instruction: |
|
Character's Persona: {bot character description} |
|
User's Persona: {user character description} |
|
Scenario: {what happens in the story} |
|
Play the role of Character. You must engage in a roleplaying chat with User below this line. Do not write dialogues and narration for User. |
|
### Input: |
|
User: {utterance} |
|
### Response: |
|
Character: {utterance} |
|
### Input: |
|
User: {utterance} |
|
### Response: |
|
Character: {utterance} |
|
(etc.) |
|
``` |
|
|
|
Inspired by the previously named "Roleplay" preset in SillyTavern, with this version of LimaRP it is possible to append a length modifier to the response instruction sequence, like this: |
|
|
|
``` |
|
### Input: |
|
User: {utterance} |
|
|
|
### Response: (length = medium) |
|
Character: {utterance} |
|
``` |
|
|
|
This has an immediately noticeable effect on bot responses. The lengths using during training are: |
|
`micro`, `tiny`, `short`, `medium`, `long`, `massive`, `huge`, `enormous`, `humongous`, `unlimited`. |
|
**The recommended starting length is medium**. Keep in mind that the AI can ramble or impersonate |
|
the user with very long messages. |
|
|
|
The length control effect is reproducible, but the messages will not necessarily follow |
|
lengths very precisely, rather follow certain ranges on average, as seen in this table |
|
with data from tests made with one reply at the beginning of the conversation: |
|
|
|
 |
|
|
|
Response length control appears to work well also deep into the conversation. **By omitting |
|
the modifier, the model will choose the most appropriate response length** (although it might |
|
not necessarily be what the user desires). |
|
|
|
## Intended uses & limitations |
|
|
|
The model will show biases similar to those observed in niche roleplaying forums on the Internet, besides those exhibited by the base model. |
|
|
|
## Training and evaluation data |
|
|
|
For more details about LimaRP, see the model page for the [previously released v2 version for Llama-2](https://huggingface.co/lemonilia/limarp-llama2-v2). |
|
|
|
## Training procedure |
|
|
|
### Training hyperparameters |
|
|
|
The following hyperparameters were used during training: |
|
- learning_rate: 0.00015 |
|
- train_batch_size: 2 |
|
- eval_batch_size: 2 |
|
- seed: 42 |
|
- gradient_accumulation_steps: 4 |
|
- total_train_batch_size: 8 |
|
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08 |
|
- lr_scheduler_type: cosine |
|
- lr_scheduler_warmup_steps: 10 |
|
- num_epochs: 2 |
|
|
|
### Training results |
|
|
|
| Training Loss | Epoch | Step | Validation Loss | |
|
|:-------------:|:-----:|:----:|:---------------:| |
|
| 2.1119 | 0.09 | 20 | 2.0727 | |
|
| 1.9889 | 0.17 | 40 | 2.0104 | |
|
| 1.878 | 0.26 | 60 | 1.9978 | |
|
| 1.8531 | 0.34 | 80 | 1.9886 | |
|
| 1.9504 | 0.43 | 100 | 1.9837 | |
|
| 1.9216 | 0.51 | 120 | 1.9826 | |
|
| 1.8483 | 0.6 | 140 | 1.9794 | |
|
| 1.9668 | 0.68 | 160 | 1.9780 | |
|
| 1.9776 | 0.77 | 180 | 1.9778 | |
|
| 1.9312 | 0.85 | 200 | 1.9772 | |
|
| 1.9003 | 0.94 | 220 | 1.9738 | |
|
| 1.8748 | 1.02 | 240 | 1.9729 | |
|
| 1.8896 | 1.11 | 260 | 1.9745 | |
|
| 1.8702 | 1.19 | 280 | 1.9760 | |
|
| 1.9038 | 1.28 | 300 | 1.9770 | |
|
| 1.9083 | 1.36 | 320 | 1.9758 | |
|
| 1.8143 | 1.45 | 340 | 1.9756 | |
|
| 1.852 | 1.53 | 360 | 1.9742 | |
|
| 1.8608 | 1.62 | 380 | 1.9735 | |
|
| 1.8959 | 1.7 | 400 | 1.9735 | |
|
| 1.7912 | 1.79 | 420 | 1.9731 | |
|
| 1.8908 | 1.87 | 440 | 1.9727 | |
|
| 1.8079 | 1.96 | 460 | 1.9729 | |
|
|
|
|
|
### Framework versions |
|
|
|
- Transformers 4.34.1 |
|
- Pytorch 2.0.1+cu118 |
|
- Datasets 2.14.6 |
|
- Tokenizers 0.14.1 |
