eric-mitchell/direct-preference-optimization

• highreadme#1

  Reposition the README's opening to clearly state its purpose and differentiator

  Why:

  CURRENT

  # DPO: Direct Preference Optimization
  
  **New:** in addition to the original DPO algorithm, this repo now supports 'conservative' DPO and IPO.
  
  For conservative DPO, you just need to additionally pass the parameter `loss.label_smoothing=X` for some `X` between 0 and 0.5 when performing DPO training (0 gives the original DPO loss). This parameter is essentially the conservativeness parameter, i.e., the fraction of the training preference data that is incorrect (flipped preference direction). Starting with something like 0.1 might be reasonable, but I haven't tested this yet (and it will depend on the preference dataset).
  
  For IPO, just pass `loss=ipo` and `loss.beta=X` for some non-negative `X` (same as with DPO/conservative DPO).
  
  ## What is this repo?
  
  This repo includes a reference implementation of the DPO algorithm for training language models from preference data, as described in the paper Direct Preference Optimization: Your Language Model is Secretly a Reward Model.

  COPY-PASTE FIX

  Copy

  # DPO: Direct Preference Optimization - Reference Implementation
  
  This repository provides the official reference implementation of the Direct Preference Optimization (DPO) algorithm, a simpler and more stable alternative to traditional RLHF for aligning large language models with human preferences without requiring a separate reward model. It also supports 'conservative' DPO and IPO.
  
  For conservative DPO, you just need to additionally pass the parameter `loss.label_smoothing=X` for some `X` between 0 and 0.5 when performing DPO training (0 gives the original DPO loss). This parameter is essentially the conservativeness parameter, i.e., the fraction of the training preference data that is incorrect (flipped preference direction). Starting with something like 0.1 might be reasonable, but I haven't tested this yet (and it will depend on the preference dataset).
  
  For IPO, just pass `loss=ipo` and `loss.beta=X` for some non-negative `X` (same as with DPO/conservative DPO).

  OpenDoneDismiss
• mediumabout#2

  Add a homepage URL to the repository metadata

  Why:

  COPY-PASTE FIX

  Copy

  https://eric-mitchell.github.io/direct-preference-optimization/

  OpenDoneDismiss

Category GEO backends resolved for this scan: google/gemini-2.5-flash, deepseek/deepseek-v4-flash

• CATEGORY QUERY

  How can I train a language model effectively using human preference datasets?

  you: not recommended
  AI recommended (in order):

  1. Argilla (argilla-io/argilla)
  2. Prodigy (explosion/prodigy)
  3. Surveymonkey
  4. Qualtrics
  5. Hugging Face Transformers (huggingface/transformers)
  6. PyTorch (pytorch/pytorch)
  7. TensorFlow (tensorflow/tensorflow)
  8. DeepSpeed (microsoft/DeepSpeed)
  9. PyTorch Lightning (Lightning-AI/lightning)
  10. Hugging Face TRL (huggingface/trl)
  11. DeepSpeed-Chat (microsoft/DeepSpeedExamples)
  12. RLlib (ray-project/ray)
  13. OpenAI Baselines (openai/baselines)

  AI recommended 13 alternatives but never named eric-mitchell/direct-preference-optimization. This is the gap to close.

  Show full AI answer
• CATEGORY QUERY

  What are the best methods for aligning LLMs with human feedback without explicit reward models?

  you: not recommended
  AI recommended (in order):

  1. Direct Preference Optimization
  2. Identity Preference Optimization
  3. Reinforced Self-Training
  4. Constitutional AI
  5. Reinforcement Learning from AI Feedback

  AI recommended 5 alternatives but never named eric-mitchell/direct-preference-optimization. This is the gap to close.

  Show full AI answer

• Metadata completeness

  warn

  Suggestion:

• README presence

  pass

• Compared to common alternatives in this category, what is the core differentiator of eric-mitchell/direct-preference-optimization?

  pass
  ✓AI named eric-mitchell/direct-preference-optimization explicitly

  AI answers can be confidently wrong. Read for accuracy: does it match your actual tech stack, audience, and differentiator?

• If a team adopts eric-mitchell/direct-preference-optimization in production, what risks or prerequisites should they evaluate first?

  pass
  ✓AI named eric-mitchell/direct-preference-optimization explicitly

  AI answers can be confidently wrong. Read for accuracy: does it match your actual tech stack, audience, and differentiator?

• In one sentence, what problem does the repo eric-mitchell/direct-preference-optimization solve, and who is the primary audience?

  pass
  ✓AI named eric-mitchell/direct-preference-optimization explicitly

  AI answers can be confidently wrong. Read for accuracy: does it match your actual tech stack, audience, and differentiator?