Merge pull request #8 from jasonvanf/codex/locate-and-fix-an-important-bug

Fix checkpoint save path for reward model

Increase the flexibility of parameters in the training reward model

Format the model training code for better readability

Added support for `resume_from_checkpoint`

Keep the maximum length consistent with 'seq_length' when setting up sft_trainer

Save model for hf trainer safely

Specifies the variable for the final checkpoint

Remove obsolete parameter layer_norm_names

Full weight fine-tuning example

Support full weight fine-tuning with DeepSpeed stage-3 (offload)

Add `warmup_ratio` and `save_total_limit` argument settings

Fix the final weights save

Support automatic merge adapter after training

Separate tokenizer setting for decapoda-research