Skip to content

v0.10.1
Compare
Choose a tag to compare
@lewtun lewtun released this 29 Aug 14:34
· 159 commits to main since this release
v0.10.1
55cc4b1

We are excited to introduce the new v0.10.1 release, with many new exciting features and post-training algorithms. The highlights are as follows:

Online DPO

Screenshot 2024-08-29 at 15 53 29

Online DPO is a new alignment method from DeepMind to boost the performance of LLMs. With Online DPO, data is generated on the fly by the trained model (instead of pre-collected). For each prompt, two completions are generated, with a reward model selecting the preferred one. This approach:

  • Eliminates the need for a pre-collected preference dataset (it's generated online)
  • Enables continuous model improvement
  • Yields better results than traditional DPO

To train models with this method, use the OnlineDPOTrainer

Liger Triton kernels for supercharged SFT

image (18)

  • We've integrated LinkedIn's Liger Triton kernels to the SFTTrainer for faster throughput and lower memory usage. To use them, set use_liger_kernel in SFTConfig

DPO for VLMs

  • We've added support to align vision-language models with DPO, now covering architectures LLaVa-1.5, PaliGemma, and Idefics2. To train VLMs with DPO, use the dpo_visual.py script as follows
accelerate launch examples/scripts/dpo_visual.py \
    --dataset_name HuggingFaceH4/rlaif-v_formatted \
    --model_name_or_path google/paligemma-3b-pt-224 \
    --trust_remote_code \
    --per_device_train_batch_size 1 \
    --gradient_accumulation_steps 8 \
    --output_dir dpo_paligemma_rlaif-v \
    --bf16 \
    --torch_dtype bfloat16

WinRate callback for LLM as a judge

  • We've added support to compute win rates over the reference model for methods like DPO. To do so, configure the callback to point to the LLM as judge API (OpenAI or Hugging Face Inference API) and then add:
trainer = DPOTrainer(...)
win_rate_callback = WinRateCallback(..., trainer=trainer)
trainer.add_callback(win_rate_callback)

Anchored Preference Optimisation (APO) for fine-grained human/AI feedback

  • Added the APO method, which is an "anchored" version of the alignment objective. There are two variants: apo_zero and apo_down. The apo_zero loss increases the likelihood of winning outputs while decreasing the likelihood of losing outputs, making it suitable when the model is less performant than the winning outputs. On the other hand, apo_down decreases the likelihood of both winning and losing outputs, but with a stronger emphasis on reducing the likelihood of losing outputs. This variant is more effective when the model is better than the winning outputs. To use these losses, set loss_type="apo_zero" or loss_type="apo_down" in the DPOConfig

What's Changed

New Contributors

Full Changelog: v0.9.6...v0.10

Contributors

kashif, kit1980, and 25 other contributors
Assets 2
Loading