qlora-yi-34b-spicyboros-3
This fine-tune took about 10 hours with single RTX 3090 Ti 24GB on Spicyboros 2.2 dataset, 0.2 epochs.
Model description
Uncensored Spicyboros based on Yi-34B, fine-tuned on local consumer GPU. I used Llama-fied Yi-34B as base, it has layer names changed to those commonly found in models with Llama architecture.
How to Use
Merge base model Yi-34B-Llama link (llama-tokenizer branch) with the adapter.bin file from this repo. Script for merging is in config folder.
Measurement.json file for Exllama v2 is in this repo too. I used spicyboros 2.2 dataset converted to parquet as a calibration dataset for the measurement.json
Intended uses & limitations
Use is limited by Yi license
Known Issues
Thesaurus mode seems to be solved by lowering repp to 1.0. I saw it going into loops of repeating sentences once or twice, but it doesn't happen all the time.
Prompt Format
ChatML prompt format is burned in. Here's a prompt format you should use
<|im_start|>system
A chat.<|im_end|>
<|im_start|>user
{prompt}<|im_end|>
<|im_start|>assistant
Training and evaluation data
I used now non-public Jon Durbin's Spicyboros 2.2 dataset. Trained on just 0.2 epochs since I already lost 4 days doing unsuccessful runs with longer training time and I got tired of that. Training config for axolotl is in this repo, it uses a preset file that is modified a little bit to fit the json formatting of the dataset I had on hand. I can share the modified preset file if you want.
Training procedure
Training hyperparameters
The following hyperparameters were used during training:
- learning_rate: 0.0001
- train_batch_size: 1
- eval_batch_size: 1
- seed: 42
- lr_scheduler_type: constant
- lr_scheduler_warmup_steps: 100
- num_epochs: 0.2
Framework versions
- Transformers 4.34.1
- Pytorch 2.0.1+cu118
- Datasets 2.14.5
- Tokenizers 0.14.1
- Downloads last month
- 3