llm-evals/llmeval/commons/admin.py

from django.contrib import admin
from django.utils.safestring import mark_safe
from django.urls import path
from .views import eval_download_results, config_download_results
from .models import (
    QA, Dataset, EvalAnswer,
    LLMBackend, LLMModel,
    EvalConfig, RoleMessage, EvalSession,
    AnswerInterpreter
    )


class QAAdmin(admin.ModelAdmin):
    list_display = ('question', 'xid', 'dataset','category', 'correct_answer', 'target')
    list_filter = ('target', 'dataset', 'category')
    search_fields = ('question', 'correct_answer', 'category', 'extra_info')

    fieldsets = (
        (None, {'fields': ('dataset', 'question', 'category', 'extra_info', 'context', 'options', 'correct_answer', 'correct_answer_idx', 'target')}),
    )

    readonly_fields = ('hash',)
admin.site.register(QA, QAAdmin)


class DatasetAdmin(admin.ModelAdmin):
    list_display = ('name', 'description')
    search_fields = ('name', 'description')
    ordering = ('-created_at',)

    fieldsets = (
        (None, {'fields': ('name', 'description')}),
    )
admin.site.register(Dataset, DatasetAdmin)


class EvalAnswerAdmin(admin.ModelAdmin):
    list_display = ('question', 'get_question_id', 'is_correct', 'llm_model', 'eval_session')
    list_filter = ('is_correct', 'llm_backend', 'llm_model', 'question__dataset', 'eval_session', 'eval_session__config')
    search_fields = ('question', 'instruction', 'assistant_answer')
    ordering = ('-created_at','question')

    def get_question_id(self, obj):
        return obj.question.id
    get_question_id.short_description = 'Question ID'
    get_question_id.admin_order_field = 'question__id'
admin.site.register(EvalAnswer, EvalAnswerAdmin)


class LLMBackendAdmin(admin.ModelAdmin):
    list_display = ('name',)
    search_fields = ('name',)
admin.site.register(LLMBackend, LLMBackendAdmin)


class LLMModelAdmin(admin.ModelAdmin):
    list_display = ('name', 'backend')
    list_filter = ('backend',)
    search_fields = ('name',)
admin.site.register(LLMModel, LLMModelAdmin)


class RoleMessageAdmin(admin.ModelAdmin):
    list_display = ('role', 'eval_config')
    list_filter = ('role', 'eval_config')
    search_fields = ('role', 'content')
admin.site.register(RoleMessage, RoleMessageAdmin)


class RoleMessageInline(admin.TabularInline):
    model = RoleMessage
    extra = 3

class EvalConfigAdmin(admin.ModelAdmin):
    list_display = ('name', 'dataset', 'created_at', 'link')
    search_fields = ('name', 'description')
    ordering = ('-created_at',)
    inlines = [RoleMessageInline]

    @admin.display(description='Link')
    def link(self, obj):
        return mark_safe(f'<a href="/commons/evalconfig/{obj.id}/config_download_results/">Download Results</a>')
    
    def get_urls(self):
        urls = super().get_urls()
        my_urls = [path('<int:config_id>/config_download_results/', config_download_results, name='config_download_results'),]
        return my_urls + urls
    
admin.site.register(EvalConfig, EvalConfigAdmin)

class EvalSessionAdmin(admin.ModelAdmin):
    list_display = ('id', 'name','is_active', 'config', 'llm_model', 'progress', 'accuracy', 'link')
    list_filter = ('is_active','config', 'llm_model')
    list_display_links = ["name"]
    search_fields = ('name', 'config', 'llm_model')
    ordering = ('-created_at',)

    @admin.display(description='Link')
    def link(self, obj):
        return mark_safe(f'<a href="/commons/evalsession/{obj.id}/eval_download_results/">Download Results</a>')
    
    def get_urls(self):
        urls = super().get_urls()
        my_urls = [path('<int:session_id>/eval_download_results/', eval_download_results, name='eval_download_results'),]
        return my_urls + urls
    
    def accuracy(self, obj):
        return "{:.2%}".format(obj.accuracy)
    
    def progress(self, obj):
        total_counts_answered = obj.evalanswer_set.count()
        total_counts = QA.objects.filter(dataset=obj.config.dataset).filter(target=obj.dataset_target).count()
        return "{}/{}".format(total_counts_answered, total_counts)

admin.site.register(EvalSession, EvalSessionAdmin)

class AnswerInterpreterAdmin(admin.ModelAdmin):
    list_display = ('name', 'llm_model')
    search_fields = ('name', 'llm_model__name')
    ordering = ('-created_at',)
admin.site.register(AnswerInterpreter, AnswerInterpreterAdmin)