eric-mitchell/direct-preference-optimization

• highreadme#1

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

  原因：

  当前

  # 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.

  复制粘贴的修复

  复制

  # 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).

  待办完成忽略
• mediumabout#2

  Add a homepage URL to the repository metadata

  原因：

  复制粘贴的修复

  复制

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

  待办完成忽略

本次扫描解析到的品类 GEO 通道：google/gemini-2.5-flash, deepseek/deepseek-v4-flash

• 品类问题

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

  你：未被推荐
  AI 推荐顺序：

  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 推荐了 13 个替代方案，却始终没点名 eric-mitchell/direct-preference-optimization。这就是要补上的差距。

  查看 AI 完整回答
• 品类问题

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

  你：未被推荐
  AI 推荐顺序：

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

  AI 推荐了 5 个替代方案，却始终没点名 eric-mitchell/direct-preference-optimization。这就是要补上的差距。

  查看 AI 完整回答

• Metadata completeness

  warn

  建议：

• README presence

  pass

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

  pass
  ✓AI 明确点名了 eric-mitchell/direct-preference-optimization

  AI 的回答可能信誓旦旦却是错的。请按事实核对：技术栈、目标人群、差异化点是不是和你实际的对得上？

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

  pass
  ✓AI 明确点名了 eric-mitchell/direct-preference-optimization

  AI 的回答可能信誓旦旦却是错的。请按事实核对：技术栈、目标人群、差异化点是不是和你实际的对得上？

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

  pass
  ✓AI 明确点名了 eric-mitchell/direct-preference-optimization

  AI 的回答可能信誓旦旦却是错的。请按事实核对：技术栈、目标人群、差异化点是不是和你实际的对得上？