Datasets:
image image |
|---|
RLHF + RLCD Rotation
First to combine RLHF + RLCD in rotation!
Idea
RLHF alone is expensive. RLCD alone is not enough.
My idea: Rotate them!
RLHF -> RLCD -> RLHF -> RLCD
How it works
- RLHF Step: Learn from human taste, human preference.
- RLCD Step: Learn from contrastive distillation, positive vs negative prompt.
- DEAD_ZONE = 0.05: If the difference between RLHF and RLCD is small, keep both ON. Don't switch. This prevents mode collapse.
- Repeat: Unknown number of steps.
Note about Code Image
The image shared is only a part of the code.
Why it works
- RLHF gives human intelligence.
- RLCD gives contrastive diversity.
- Rotation prevents repetition.
- DEAD_ZONE prevents collapse.
RLHF + RLCD = 0.7, 0.8, 0.9. Real thinking.
Results
- No mode collapse.
- No repetition.
- Model stays smart, not dumber.
Training
This training is not finished yet. It needs more resources to reach the target.
Setup
- Base Model: There is no hope yet of reaching 100 million, but it is expandable without problem.
- Method: RLHF + RLCD Rotation
- LoRA: Yes, optimized LoRA
- dtype: float32
- Optimizer: AdamW (lr=2e-4)
Why expandable?
Because it contains an idea like a person who reads a book once and forgets it ππ But it stays stored in his mind, and he returns to read it many times to memorize it. That is real training.
Why it needs more resources?
With LoRA + float32, training is heavier.
- float32 keeps weights precise, no loss.
- LoRA needs optimization.
- group_params is used for the adapter only, for optimization, nothing else.
This costs more RAM / VRAM, but it is necessary to get the correct result.
group_params
I made an improvement for LoRA:
group_paramsis only for the adapter.- Only adapter weights are optimized.
- Base model is frozen.
- No conflict at all.
Checked
- Full code checked.
- Adapter checked.
- Weights checked.
- No conflict at all!
What is needed?
To finish training and get the required result, we need more resources:
- More VRAM / RAM
- Clean run, no interruption
Status
Code is clean. Code is complete. Code is checked. Training is waiting for resources.
Code
No code shared. Idea only.
Author
Mouad Tarif
Contact
email:mouad.tarif@inbox.ru
linkedin:https://www.linkedin.com/in/mouad-tarif-297688340
**Important note**
I thought about a training with a completely different code structure than that one π
And a small hint: those who used Fisher, which is a well-known function from 2017, I used the same method, but without following those who made it.
The important thing is, those who worked on the training made it so weights don't change because of a penalty, and that led to forgetting.
I hope for help and support for what I worked on. I don't think so at all, so I will find different ideas for training on the resources I have.
I opened a new communication channel, a new official account on LinkedIn, and I put the link to my personal account.
This idea I published about RLCD and RLHF is from my own thinking, officially, not from anyone else. And here are the different ideas I have.
- Downloads last month
- 91
