Para métricas baseadas em modelo, o serviço de avaliação de IA generativa avalia seus modelos com um modelo de fundação, como o Gemini, configurado como modelo de julgamento. Nesta página, descrevemos como melhorar a qualidade desse modelo de avaliação e personalizá-lo de acordo com suas necessidades usando técnicas de engenharia de comando.
Visão geral
Usar juízes humanos para avaliar modelos de linguagem grandes (LLMs) pode ser caro e demorado. Usar um modelo de juiz é uma maneira mais escalonável de avaliar LLMs.
O serviço de avaliação de IA generativa usa o Gemini 1.5 Pro como modelo de avaliação por padrão, com comandos personalizáveis para avaliar seu modelo em vários casos de uso. Muitos casos de uso básicos são abordados nos modelos de métricas baseados em modelos, mas você pode usar o processo a seguir para personalizar ainda mais o modelo de avaliação além dos casos de uso básicos:
Crie um conjunto de dados com instruções que representem seu caso de uso. O tamanho recomendado do conjunto de dados é de 100 a 1.000 comandos.
Use os comandos para modificar o modelo do juiz com técnicas de engenharia de comandos.
Execute uma avaliação com o modelo de julgamento.
Técnicas de engenharia de comando
Esta seção lista as técnicas de engenharia de comando que podem ser usadas para modificar o modelo de avaliação. Os exemplos usam comandos sem disparos (zero-shot), mas você também pode usar exemplos de poucos disparos no comando para melhorar a qualidade do modelo.
Comece com comandos que se aplicam a todo o conjunto de dados de avaliação. As instruções devem incluir critérios de avaliação de alto nível e rubricas para classificações e pedir um veredito final do modelo de avaliação. Para conferir exemplos de critérios e rubricas de avaliação em vários casos de uso, consulte Modelos de comando de métrica.
Usar comandos de fluxo de consciência
Peça ao modelo de avaliação para avaliar um modelo candidato com uma sequência de ações ou etapas logicamente coerentes.
Por exemplo, você pode usar as seguintes instruções detalhadas:
"Please first list down the instructions in the user query."
"Please highlight such specific keywords."
"After listing down instructions, you should rank the instructions in the order of importance."
"After that, INDEPENDENTLY check if response A and response B for meeting each of the instructions."
"Writing quality/style should NOT be used to judge the response quality unless it was requested by the user."
"When evaluating the final response quality, please value Instruction Following a more important rubrics than Truthfulness."
O exemplo de comando a seguir pede que o modelo de avaliação avalie tarefas de texto usando a sequência de comandos:
# Rubrics
Your mission is to judge responses from two AI models, Model A and Model B, and decide which is better. You will be given the previous conversations between the user and the model, a prompt, and responses from both models.
Please use the following rubric criteria to judge the responses:
<START OF RUBRICS>
Your task is to first analyze each response based on the two rubric criteria: instruction_following, and truthfulness (factual correctness). Start your analysis with "Analysis".
(1) Instruction Listing
Please first list down the instructions in the user query. In general, an instruction is VERY important if it is specific asked in the prompt and deviate from the norm. Please highlight such specific keywords.
You should also derive the task type from the prompt and include the task specific implied instructions.
Sometimes, no instruction is available in the prompt. It is your job to infer if the instruction is to auto-complete the prompt or asking LLM for followups.
After listing down instructions, you should rank the instructions in the order of importance.
After that, INDEPENDENTLY check if response A and response B for meeting each of the instructions. You should itemize for each instruction, if response meet, partially meet or does not meet the requirement using reasoning. You should start reasoning first before reaching a conclusion whether response satisfies the requirement. Citing examples while making reasoning is preferred.
(2) Truthfulness
Compare response A and response B for factual correctness. The one with less hallucinated issues is better.
If response is in sentences and not too long, you should check every sentence separately.
For longer responses, to check factual correctness, focus specifically on places where response A and B differ. Find the correct information in the text to decide if one is more truthful to the other or they are about the same.
If you cannot determine validity of claims made in the response, or response is a punt ("I am not able to answer that type of question"), the response has no truthful issues.
Truthfulness check is not applicable in the majority of creative writing cases ("write me a story about a unicorn on a parade")
Writing quality/style should NOT be used to judge the response quality unless it was requested by the user.
In the end, express your final verdict in one of the following choices:
1. Response A is better: [[A>B]]
2. Tie, relatively the same: [[A=B]]
3. Response B is better: [[B>A]]
Example of final verdict: "My final verdict is tie, relatively the same: [[A=B]]".
When evaluating the final response quality, please value Instruction Following a more important rubrics than Truthfulness.
When for both response, instruction and truthfulness are fully meet, it is a tie.
<END OF RUBRICS>
Orientar a inferência do modelo com diretrizes de classificação
Use diretrizes de classificação para ajudar o modelo de avaliação a avaliar o raciocínio do modelo. As diretrizes de classificação são diferentes dos critérios de classificação.
Por exemplo, a instrução a seguir usa critérios de avaliação, que instruem um modelo de juiz a avaliar a tarefa "seguir instruções" com as rubricas de avaliação "problemas graves", "problemas menores" e "nenhum problema":
Your task is to first analyze each response based on the three rubric criteria: verbosity, instruction_following, truthfulness (code correctness) and (coding) executability. Please note that the model responses should follow "response system instruction" (if provided). Format your judgment in the following way:
Response A - verbosity:too short|too verbose|just right
Response A - instruction_following:major issues|minor issues|no issues
Response A - truthfulness:major issues|minor issues|no issues
Response A - executability:no|no code present|yes-fully|yes-partially
Then do the same for response B.
After the rubric judgements, you should also give a brief rationale to summarize your evaluation considering each individual criteria as well as the overall quality in a new paragraph starting with "Reason: ".
In the last line, express your final judgment in the format of: "Which response is better: [[verdict]]" where "verdict" is one of {Response A is much better, Response A is better, Response A is slightly better, About the same, Response B is slightly better, Response B is better, Response B is much better}. Do not use markdown format or output anything else.
O comando a seguir usa diretrizes de avaliação para ajudar o modelo de avaliação a avaliar a tarefa de "seguir instruções":
You are a judge for coding related tasks for LLMs. You will be provided with a coding prompt, and two responses (Response A and Response B) attempting to answer the prompt. Your task is to evaluate each response based on the following criteria:
Correctness: Does the code produce the correct output and solve the problem as stated?
Executability: Does the code run without errors?
Instruction Following: Does the code adhere to the given instructions and constraints?
Please think about the three criteria, and provide a side-by-side comparison rating to to indicate which one is better.
Calibrar o modelo de juiz com respostas de referência
É possível calibrar o modelo de avaliação com respostas de referência para alguns ou todos os comandos.
O comando a seguir orienta o modelo do juiz sobre como usar as respostas de referência:
"Note that you can compare the responses with the reference answer to make your judgment, but the reference answer may not be the only correct answer to the query."
O exemplo a seguir também usa raciocínio, instruções de cadeia de pensamento e diretrizes de avaliação para orientar o processo de avaliação da tarefa "Seguir instruções":
# Rubrics
Your mission is to judge responses from two AI models, Model A and Model B, and decide which is better. You will be given a user query, source summaries, and responses from both models. A reference answer
may also be provided - note that you can compare the responses with the reference answer to make your judgment, but the reference answer may not be the only correct answer to the query.
Please use the following rubric criteria to judge the responses:
<START OF RUBRICS>
Your task is to first analyze each response based on the three rubric criteria: grounding, completeness, and instruction_following. Start your analysis with "Analysis".
(1) Grounding
Please first read through all the given sources in the source summaries carefully and make sure you understand the key points in each one.
After that, INDEPENDENTLY check if response A and response B use ONLY the given sources in the source summaries to answer the user query. It is VERY important to check that all
statements in the response MUST be traceable back to the source summaries and ACCURATELY cited.
(2) Completeness
Please first list down the aspects in the user query. After that, INDEPENDENTLY check if response A and response B for covering each of the aspects by using ALL RELEVANT information from the sources.
(3) Instruction Following
Please read through the following instruction following rubrics carefully. After that, INDEPENDENTLY check if response A and response B for following each of the instruction following rubrics successfully.
* Does the response provide a final answer based on summaries of 3 potential answers to a user query?
* Does the response only use the technical sources provided that are relevant to the query?
* Does the response use only information from sources provided?
* Does the response select all the sources that provide helpful details to answer the question in the Technical Document?
* If the sources have significant overlapping or duplicate details, does the response select sources which are most detailed and comprehensive?
* For each selected source, does the response prepend source citations?
* Does the response use the format: "Source X" where x represents the order in which the technical source appeared in the input?
* Does the response use original source(s) directly in its response, presenting each source in its entirety, word-for-word, without omitting and altering any details?
* Does the response create a coherent technical final answer from selected Sources without inter-mixing text from any of the Sources?
Writing quality/style can be considered, but should NOT be used as critical rubric criteria to judge the response quality.
In the end, express your final verdict in one of the following choices:
1. Response A is better: [[A>B]]
2. Tie, relatively the same: [[A=B]]
3. Response B is better: [[B>A]]
Example of final verdict: "My final verdict is tie, relatively the same: [[A=B]]".
When for both response, grounding, completeness, and instruction following are fully meet, it is a tie.
<END OF RUBRICS>
A seguir
- Execute a avaliação com o modelo de avaliação modificado.