yujunjun/LLMs-from-scratch
1
0
Fork 0
mirror of https://github.com/rasbt/LLMs-from-scratch.git synced 2025-08-08 00:32:11 +00:00
Code Issues Packages Projects Releases Wiki Activity

add more explanations

Browse Source
This commit is contained in:
rasbt 2024-08-06 19:45:11 -05:00
parent 98d24a1607
commit c8090f30ef
No known key found for this signature in database
GPG Key ID: 3C6E5C7C075611DB
1 changed files with 5 additions and 4 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

9
ch07/04_preference-tuning-with-dpo/dpo-from-scratch.ipynb
View File

@ -154,10 +154,11 @@
},
"source": [
"- In the equation above,\n",
" - \"expected value\" $\\mathbb{E}$ is statistics jargon and stands for the average or mean value of the random variable (the expression inside the brackets)\n",
" - The $\\pi_{\\theta}$ variable is the so-called policy (a term borrowed from reinforcement learning) and represents the LLM we want to optimize; $\\pi_{ref}$ is a reference LLM, which is typically the original LLM before optimization (at the beginning of the training, $\\pi_{\\theta}$ and $\\pi_{ref}$ are typically the same)\n",
" - $\\beta$ is a hyperparameter to control the divergence between the $\\pi_{\\theta}$ and the reference model; increasing $\\beta$ increases the impact of the difference between\n",
" - \"expected value\" $\\mathbb{E}$ is statistics jargon and stands for the average or mean value of the random variable (the expression inside the brackets); optimizing $-\\mathbb{E}$ aligns the model better with user preferences\n",
" - The $\\pi_{\\theta}$ variable is the so-called policy (a term borrowed from reinforcement learning) and represents the LLM we want to optimize; $\\pi_{ref}$ is a reference LLM, which is typically the original LLM before optimization (at the beginning of the training, $\\pi_{\\theta}$ and $\\pi_{ref}$ are typically the same)\n",
" - $\\beta$ is a hyperparameter to control the divergence between the $\\pi_{\\theta}$ and the reference model; increasing $\\beta$ increases the impact of the difference between\n",
"$\\pi_{\\theta}$ and $\\pi_{ref}$ in terms of their log probabilities on the overall loss function, thereby increasing the divergence between the two models\n",
" - the logistic sigmoid function, $\\log \\sigma(\\centerdot)$ transforms the log-odds of the preferred and rejected responses (the terms inside the logistic sigmoid function) into a log-probability score \n",
"- To avoid bloating the code notebook with a more detailed discussion, I may write a separate standalone article with more details on these concepts in the future\n",
"- In the meantime, if you are interested in comparing RLHF and DPO, please see the section [2.2. RLHF vs Direct Preference Optimization (DPO)](https://magazine.sebastianraschka.com/i/142924793/rlhf-vs-direct-preference-optimization-dpo) in my article [Tips for LLM Pretraining and Evaluating Reward Models](https://magazine.sebastianraschka.com/p/tips-for-llm-pretraining-and-evaluating-rms)"
]
@ -3088,7 +3089,7 @@
"name": "python",
"nbconvert_exporter": "python",
"pygments_lexer": "ipython3",
"version": "3.10.6"
"version": "3.11.4"
}
},
"nbformat": 4,