metadata
license: apache-2.0
base_model: one-man-army/una-neural-chat-v3-3-P1-OMA
tags:
- alignment-handbook
- generated_from_trainer
datasets:
- allenai/ultrafeedback_binarized_cleaned
model-index:
- name: una-neural-chat-v3-3-P2
results: []
OMA, OneManArmy presents, una-neural-chat-v3-3 PHASE 2. Powered by UNA (Uniform Neural Alignment), using zephyr trainer, allenai/ultrafeedback cleaned.. and JUST THAT.
Outperforming its base model, not adding any data.. just UNA Algorythm on Transformers Lib.
UNA Settings:
- MLP : 0.05
- ATT : 0.03
- LNOR : 0.02
una-neural-chat-v3-3-phase2
This model is a fine-tuned version of Intel/neural-chat-7b-v3-3 on the allenai/ultrafeedback_binarized_cleaned dataset. It achieves the following results on the evaluation set:
- Loss: 0.4524
- Rewards/chosen: -0.7101
- Rewards/rejected: -2.0953
- Rewards/accuracies: 0.7831
- Rewards/margins: 1.3852
- Logps/rejected: -321.5471
- Logps/chosen: -327.5048
- Logits/rejected: -2.6445
- Logits/chosen: -2.6674
Training results
| Training Loss | Epoch | Step | Validation Loss | Rewards/chosen | Rewards/rejected | Rewards/accuracies | Rewards/margins | Logps/rejected | Logps/chosen | Logits/rejected | Logits/chosen |
|---|---|---|---|---|---|---|---|---|---|---|---|
| 0.5431 | 0.2 | 380 | 0.4900 | -0.6823 | -1.6613 | 0.7607 | 0.9790 | -317.2069 | -327.2263 | -2.6478 | -2.6651 |
| 0.4369 | 0.4 | 760 | 0.4783 | -0.7562 | -2.1298 | 0.7719 | 1.3737 | -321.8924 | -327.9652 | -2.7370 | -2.7562 |
| 0.4005 | 0.6 | 1140 | 0.4697 | -0.6913 | -2.0134 | 0.7770 | 1.3221 | -320.7278 | -327.3167 | -2.7067 | -2.7224 |
| 0.3759 | 0.8 | 1520 | 0.4568 | -0.7387 | -2.0643 | 0.7882 | 1.3256 | -321.2370 | -327.7909 | -2.6626 | -2.6829 |
| 0.5213 | 1.0 | 1900 | 0.4524 | -0.7101 | -2.0953 | 0.7831 | 1.3852 | -321.5471 | -327.5048 | -2.6445 | -2.6674 |
Framework versions
- Transformers 4.35.0-UNA
- Pytorch 2.1.0
- Datasets 2.14.6
- Tokenizers 0.14.1