Circular import issue

[kekePower]

Copied from LeapLabTHU/Absolute-Zero-Reasoner#7 (comment)

Today I have been trying to use a GROQ LLM (namely: "llama-3.3-70b-versatile", because "mixtral-8x7b-32768" is DEPRECATED) for all 3 functions:
Proposer/Solver/Evaluator

In config\llm_config.yaml :
proposer: llama-3.3-70b-versatile # Instead of OpenAI's smaller model for proposal generation
solver: llama-3.3-70b-versatile # Instead of Qwen3 for solving complex tasks
evaluator: llama-3.3-70b-versatile # Instead of DeepSeek for evaluation

Also I followed your instructions in the README.md file to get started.

Clone the repository
Copy config/llm_config.yaml.example to config/llm_config.yaml

In llm_config.yaml I changes these things:
default_provider: openai

default_provider: groq

proposer: llama-3.3-70b-versatile # Instead of OpenAI's smaller model for proposal generation
solver: llama-3.3-70b-versatile # Instead of Qwen3 for solving complex tasks
evaluator: llama-3.3-70b-versatile # Instead of DeepSeek for evaluation
Model-specific settings

model_settings:

"llama-3.3-70b-versatile"
temperature: 0.5
max_tokens: 4000
top_p: 0.9

Set your API keys as environment variables

Here I made a .env file:
GROQ_API_KEY="gsk_......"

Run python azr_ukg.py

But no matter what I tried, I keep getting this error trace (I use CURSOR as my IDE):

F:\Downloads\absolute-zero-universal-knowledge-main\absolute-zero-universal-knowledge-features-multi-llm-support>python azr_ukg.py
Traceback (most recent call last):
File "F:\Downloads\absolute-zero-universal-knowledge-main\absolute-zero-universal-knowledge-features-multi-llm-support\azr_ukg.py", line 18, in
from modules.config import (
File "F:\Downloads\absolute-zero-universal-knowledge-main\absolute-zero-universal-knowledge-features-multi-llm-support\modules\config_init_.py", line 7, in
from .loader import (
File "F:\Downloads\absolute-zero-universal-knowledge-main\absolute-zero-universal-knowledge-features-multi-llm-support\modules\config\loader.py", line 12, in
from ..llm.rate_limit import RateLimitConfig
File "F:\Downloads\absolute-zero-universal-knowledge-main\absolute-zero-universal-knowledge-features-multi-llm-support\modules\llm_init_.py", line 12, in
from .providers.base import (
File "F:\Downloads\absolute-zero-universal-knowledge-main\absolute-zero-universal-knowledge-features-multi-llm-support\modules\llm\providers_init_.py", line 24, in
from . import openai
File "F:\Downloads\absolute-zero-universal-knowledge-main\absolute-zero-universal-knowledge-features-multi-llm-support\modules\llm\providers\openai.py", line 67, in
from .. import LLMFactory
ImportError: cannot import name 'LLMFactory' from partially initialized module 'modules.llm' (most likely due to a circular import) (F:\Downloads\absolute-zero-universal-knowledge-main\absolute-zero-universal-knowledge-features-multi-llm-support\modules\llm_init_.py)
F:\Downloads\absolute-zero-universal-knowledge-main\absolute-zero-universal-knowledge-features-multi-llm-support>

Could this be caused by a circular import?
Have I forgotten something?
And most importantly: What can be the solution to this?

[kekePower]

I've just pushed an update that I believe fixes the circular import issues.
Make a backup of your configuration, do a new test and let me know.

[jbdatascience]

I've just pushed an update that I believe fixes the circular import issues. Make a backup of your configuration, do a new test and let me know.

We got rid of the circular import issues!

But a problem remains: In my first experiments I want to use GROQ model "llama-3.3-70b-versatile" for all LLM instances in your code (so for the Proposer/Solver/Critiquer/Eval1 etc. and also for the OPENAI_QUESTION_MODEL !).

But at the moment it does not use the GROQ model "llama-3.3-70b-versatile" for the Question Model (but it tries to use OpenAI Proposer LLM (gpt-4.1-mini). I do not have an OpenAI account, so I must resort to a GROQ model for this!

Can you show exactly what I do have to change in the code so that only GROQ model "llama-3.3-70b-versatile" is used for all LLMs used in the code?

A small example of the output I get when running:
python azr_ukg.py

(

It says:

Primary Model (Proposer/Solver/Critiquer/Eval1): 'llama-3.3-70b-versatile' via https://api.novita.ai/v3/openai

And

Secondary Evaluator Model (Eval2): deepseek/deepseek-v3-0324 (uses Primary API Base URL &
Key)

Although I tried to configure GROQ model "llama-3.3-70b-versatile" for all MODEL uses in the code, together with its API key and API_BASE_URL, but I must have something wrong !

)

OUTPUT:

---

DISCLAIMER: This script generates highly speculative and abstract content using LLMs.
The generated tasks and solutions are for research and exploration into AI capabilities.
Interpret outputs with caution; they are not validated facts or established knowledge.
Ensure API Keys (e.g., PRIMARY_API_KEY) are set or updated.
The 'openai' library is required (pip install --upgrade openai for AsyncOpenAI).

---

OpenAI library version: 1.79.0
Starting Absolute Zero Universal Knowledge Generator (v1.4.7 - Panel, Self-Critique, Perturbations)...
Primary Model (Proposer/Solver/Critiquer/Eval1): 'llama-3.3-70b-versatile' via https://api.novita.ai/v3/openai
Secondary Evaluator Model (Eval2): deepseek/deepseek-v3-0324 (uses Primary API Base URL &
Key)
Logging explorations to: universal_knowledge_exploration_log_v1.4.7.jsonl
Targeting API RPM Limit: 100
Solver rollouts for proposer reward: 1
Logging Quality Threshold: 0.3, Learned Concept Threshold: 0.6
Stochastic Perturbation Probability for Proposer: 15.0%

--- Iteration 1/30 ---
Selected task type for proposer: first_principles_reimagination
Concept Proposer: Attempting to generate a new seed concept.
Applying stochastic seed to proposer: 'novel frameworks for interdisciplinary knowledge
synthesis'

DEBUG: Final prompt for OpenAI Proposer LLM (gpt-4.1-mini):

You are an AI assistant tasked with generating a new, challenging, and unique task of the
type: 'first_principles_reimagination'.
Your goal is to propose a task that is novel and has clearly defined success criteria.

<<<TASK_SPECIFIC_REFINEMENT_AREA>>>

Follow these structural guidelines meticulously:

1. Think Step-by-Step: First, engage in a brief block to outline your thought process for creating the task. This should include considerations for novelty, clarity of success criteria, and adherence to the task type. State the target novelty level you are aiming for (0.0 to 1.0).
2. Provide JSON Output: Immediately after the closing tag, provide your response as a single, valid JSON object.

JSON Structure Requirements:

• "task_description": A detailed description of the task.
• "success_criteria": Specific, measurable criteria for evaluating a successful solution (as a string).
• "domain_tags": A list of 2-5 relevant domain tags (e.g., ['physics', 'philosophy']).
• "novelty_level": An estimated novelty score from 0.0 (mundane) to 1.0 (paradigm-shifting).
• "task_type_generated": Echo back the precise task type you were asked to generate (e.g., 'first_principles_reimagination').

CRITICAL: The JSON object MUST immediately follow the closing tag. It MUST be a raw JSON string. DO NOT wrap it in Markdown code blocks (e.g., json ... ). The JSON must start with { and end with }.
Ensure that all string values within the JSON (especially 'task_description' and 'success_criteria') are plain text and contain NO MARKDOWN formatting.

Here's an example of the complete output structure (content is illustrative):

I will design a 'first_principles_reimagination' task. My aim is a novelty level of 0.7. I need to ensure the success criteria are very specific and test for deep understanding. The domain tags should be relevant to the problem's core.
{
"task_description": "(A detailed description of the novel task, explicitly related to 'first_principles_reimagination')...",
"success_criteria": "(Clear, measurable, and unambiguous criteria for evaluating a solution. This should detail what a successful response must include and how it will be judged)...",
"domain_tags": ["relevant_tag_1", "relevant_tag_2", "interdisciplinary_tag"],
"novelty_level": 0.7,
"task_type_generated": "(Must be exactly 'first_principles_reimagination')"
}

---

Sending request to OpenAI model: gpt-4.1-mini (Temp: 0.9, Max Tokens: 1000)
OpenAI System Prompt: You are a helpful AI assistant that follows instructions precisely and returns outputs in the reques...
OpenAI User Prompt: You are an AI assistant tasked with generating a new, challenging, and unique task of the type: 'first_principles_reimagination'.
Your goal is to prop...
OpenAI APIStatusError: 401 - <Response [401 Unauthorized]>
DEBUG: Raw Proposer (OpenAI) response:
None

ERROR: Proposer LLM (OpenAI) failed to generate a response. Skipping this iteration.

--- Iteration 2/30 ---
Selected task type for proposer: panel_discussion_challenge
Concept Proposer: Attempting to generate a new seed concept.

DEBUG: Final prompt for OpenAI Proposer LLM (gpt-4.1-mini):

You are an AI assistant tasked with generating a new, challenging, and unique task of the
type: 'panel_discussion_challenge'.
Your goal is to propose a task that is novel and has clearly defined success criteria.

<<<TASK_SPECIFIC_REFINEMENT_AREA>>>

Follow these structural guidelines meticulously:

1. Think Step-by-Step: First, engage in a brief block to outline your thought process for creating the task. This should include considerations for novelty, clarity of success criteria, and adherence to the task type. State the target novelty level you are aiming for (0.0 to 1.0).
2. Provide JSON Output: Immediately after the closing tag, provide your response as a single, valid JSON object.

JSON Structure Requirements:

• "task_description": A detailed description of the task.
• "success_criteria": Specific, measurable criteria for evaluating a successful solution (as a string).
• "domain_tags": A list of 2-5 relevant domain tags (e.g., ['physics', 'philosophy']).
• "novelty_level": An estimated novelty score from 0.0 (mundane) to 1.0 (paradigm-shifting).
• "task_type_generated": Echo back the precise task type you were asked to generate (e.g., 'panel_discussion_challenge').

CRITICAL: The JSON object MUST immediately follow the closing tag. It MUST be a raw JSON string. DO NOT wrap it in Markdown code blocks (e.g., json ... ). The JSON must start with { and end with }.
Ensure that all string values within the JSON (especially 'task_description' and 'success_criteria') are plain text and contain NO MARKDOWN formatting.

Here's an example of the complete output structure (content is illustrative):

I will design a 'panel_discussion_challenge' task. My aim is a novelty level of 0.7. I need to ensure the success criteria are very specific and test for deep understanding. The domain tags should be relevant to the problem's core.
{
"task_description": "(A detailed description of the novel task, explicitly related to 'panel_discussion_challenge')...",
"success_criteria": "(Clear, measurable, and unambiguous criteria for evaluating a solution. This should detail what a successful response must include and how it will be judged)...",
"domain_tags": ["relevant_tag_1", "relevant_tag_2", "interdisciplinary_tag"],
"novelty_level": 0.7,
"task_type_generated": "(Must be exactly 'panel_discussion_challenge')"
}

---

Sending request to OpenAI model: gpt-4.1-mini (Temp: 0.9, Max Tokens: 1000)
OpenAI System Prompt: You are a helpful AI assistant that follows instructions precisely and returns outputs in the reques...
OpenAI User Prompt: You are an AI assistant tasked with generating a new, challenging, and unique task of the type: 'panel_discussion_challenge'.
Your goal is to propose ...
OpenAI APIStatusError: 401 - <Response [401 Unauthorized]>
DEBUG: Raw Proposer (OpenAI) response:
None

ERROR: Proposer LLM (OpenAI) failed to generate a response. Skipping this iteration.

--- Iteration 3/30 ---
Selected task type for proposer: generation_of_novel_axioms_and_exploration
Concept Proposer: Attempting to generate a new seed concept.

DEBUG: Final prompt for OpenAI Proposer LLM (gpt-4.1-mini):

You are an AI assistant tasked with generating a new, challenging, and unique task of the
type: 'generation_of_novel_axioms_and_exploration'.
Your goal is to propose a task that is novel and has clearly defined success criteria.

<<<TASK_SPECIFIC_REFINEMENT_AREA>>>

Follow these structural guidelines meticulously:

1. Think Step-by-Step: First, engage in a brief block to outline your thought process for creating the task. This should include considerations for novelty, clarity of success criteria, and adherence to the task type. State the target novelty level you are aiming for (0.0 to 1.0).
2. Provide JSON Output: Immediately after the closing tag, provide your response as a single, valid JSON object.

JSON Structure Requirements:

• "task_description": A detailed description of the task.
• "success_criteria": Specific, measurable criteria for evaluating a successful solution (as a string).
• "domain_tags": A list of 2-5 relevant domain tags (e.g., ['physics', 'philosophy']).
• "novelty_level": An estimated novelty score from 0.0 (mundane) to 1.0 (paradigm-shifting).
• "task_type_generated": Echo back the precise task type you were asked to generate (e.g., 'generation_of_novel_axioms_and_exploration').

CRITICAL: The JSON object MUST immediately follow the closing tag. It MUST be a raw JSON string. DO NOT wrap it in Markdown code blocks (e.g., json ... ). The JSON must start with { and end with }.
Ensure that all string values within the JSON (especially 'task_description' and 'success_criteria') are plain text and contain NO MARKDOWN formatting.

Here's an example of the complete output structure (content is illustrative):

I will design a 'generation_of_novel_axioms_and_exploration' task. My aim is a novelty level of 0.7. I need to ensure the success criteria are very specific and test for deep understanding. The domain tags should be relevant to the problem's core.
{
"task_description": "(A detailed description of the novel task, explicitly related to 'generation_of_novel_axioms_and_exploration')...",
"success_criteria": "(Clear, measurable, and unambiguous criteria for evaluating a solution. This should detail what a successful response must include and how it will be judged)...",
"domain_tags": ["relevant_tag_1", "relevant_tag_2", "interdisciplinary_tag"],
"novelty_level": 0.7,
"task_type_generated": "(Must be exactly 'generation_of_novel_axioms_and_exploration')"
}

---

Sending request to OpenAI model: gpt-4.1-mini (Temp: 0.9, Max Tokens: 1000)
OpenAI System Prompt: You are a helpful AI assistant that follows instructions precisely and returns outputs in the reques...
OpenAI User Prompt: You are an AI assistant tasked with generating a new, challenging, and unique task of the type: 'generation_of_novel_axioms_and_exploration'.
Your goa...
OpenAI APIStatusError: 401 - <Response [401 Unauthorized]>
DEBUG: Raw Proposer (OpenAI) response:
None

ERROR: Proposer LLM (OpenAI) failed to generate a response. Skipping this iteration.
Traceback (most recent call last):
File "C:\ProgramData\anaconda3\Lib\asyncio\runners.py", line 118, in run
return self._loop.run_until_complete(task)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\ProgramData\anaconda3\Lib\asyncio\base_events.py", line 653, in run_until_complete
return future.result()
^^^^^^^^^^^^^^^
File "F:\Downloads\absolute-zero-universal-knowledge-main\absolute-zero-universal-knowledge-features-multi-llm-support (1)\absolute-zero-universal-knowledge-features-multi-llm-support\azr_ukg.py", line 255, in main
await asyncio.sleep(MIN_ITER_SLEEP * 5)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "C:\ProgramData\anaconda3\Lib\asyncio\tasks.py", line 649, in sleep
return await future
^^^^^^^^^^^^
asyncio.exceptions.CancelledError

During handling of the above exception, another exception occurred:

Traceback (most recent call last):
File "F:\Downloads\absolute-zero-universal-knowledge-main\absolute-zero-universal-knowledge-features-multi-llm-support (1)\absolute-zero-universal-knowledge-features-multi-llm-support\azr_ukg.py", line 486, in
asyncio.run(main())
File "C:\ProgramData\anaconda3\Lib\asyncio\runners.py", line 190, in run
return runner.run(main)
^^^^^^^^^^^^^^^^
File "C:\ProgramData\anaconda3\Lib\asyncio\runners.py", line 123, in run
raise KeyboardInterrupt()
KeyboardInterrupt
^C
(base) F:\Downloads\absolute-zero-universal-knowledge-main\absolute-zero-universal-knowledge-features-multi-llm-support (1)\absolute-zero-universal-knowledge-features-multi-llm-support>

I would appreciate your help with overcoming these start problems !

PS: I am working on Windows

[kekePower]

Thanks. Will look into it. Have a bit on the table at the moment, so it may take a while.

[jbdatascience]

Thanks. Will look into it. Have a bit on the table at the moment, so it may take a while.

No problem.

In the meantime I am very curious about the results you had already with AZR-UKG !