Agent Training Code

[JadeCityYC]

Dear author @cz3716087,

Thanks for your wonderful work.

I just want to confirm if there is no code for training the deep research agentic system.

Are the open_data.jsonl you mentioned the training dataset for training the agentic system? And are the rewritten_result.jsonl you mentioned the corresponding training dataset we can generate by the Agentic Dataset Construction module? If we follow the same settings with you, the rewritten_result.jsonl should be the same as open_data.jsonl, am I right?

Hope to get the response from your side. Thanks!

[cz3716087]

Dear author @cz3716087,

Thanks for your wonderful work.

I just want to confirm if there is no code for training the deep research agentic system.

Are the open_data.jsonl you mentioned the training dataset for training the agentic system? And are the rewritten_result.jsonl you mentioned the corresponding training dataset we can generate by the Agentic Dataset Construction module? If we follow the same settings with you, the rewritten_result.jsonl should be the same as open_data.jsonl, am I right?

Hope to get the response from your side. Thanks!

Hi @JadeCityYC,

Thank you for your interest in the project.

Regarding the training code for the deep research agentic system, we haven't provided dedicated scripts in the repository. For Supervised Fine-Tuning (SFT), you can use frameworks like Megatron or LLaMA-Factory and proceed as you would with training other LLM models. For Reinforcement Learning (RL), it's a bit more involved, as you'll need to set up the tool environment yourself (which isn't included in our system). You might want to explore libraries like VERL to build it from scratch.

As for the datasets: open_data.jsonl is generated from the KnowledgeGraphConstruction module and serves as the input for trajectory generation. You'll then need to use your preferred model with the TrajectoryGenerationPipeline to create the trajectories, which will produce rewritten_result.jsonl as the output. Therefore, open_data.jsonl is the input data used to generate rewritten_result.jsonl.

If you have any further questions, feel free to ask!

[JadeCityYC]

Dear author @cz3716087,
Thanks for your wonderful work.
I just want to confirm if there is no code for training the deep research agentic system.
Are the open_data.jsonl you mentioned the training dataset for training the agentic system? And are the rewritten_result.jsonl you mentioned the corresponding training dataset we can generate by the Agentic Dataset Construction module? If we follow the same settings with you, the rewritten_result.jsonl should be the same as open_data.jsonl, am I right?
Hope to get the response from your side. Thanks!

Hi @JadeCityYC,

Thank you for your interest in the project.

Regarding the training code for the deep research agentic system, we haven't provided dedicated scripts in the repository. For Supervised Fine-Tuning (SFT), you can use frameworks like Megatron or LLaMA-Factory and proceed as you would with training other LLM models. For Reinforcement Learning (RL), it's a bit more involved, as you'll need to set up the tool environment yourself (which isn't included in our system). You might want to explore libraries like VERL to build it from scratch.

As for the datasets: open_data.jsonl is generated from the KnowledgeGraphConstruction module and serves as the input for trajectory generation. You'll then need to use your preferred model with the TrajectoryGenerationPipeline to create the trajectories, which will produce rewritten_result.jsonl as the output. Therefore, open_data.jsonl is the input data used to generate rewritten_result.jsonl.

If you have any further questions, feel free to ask!

Dear author @cz3716087,

Thanks for your quick response.

Based on your response, so rewritten_result.jsonl is the training dataset for the deep research agentic system, am I right?

Because you have released the pre-trained 32B model in HuggingFace, does that mean I can load the model and fine-tune the parameters of the agentic system using your existing code?

I am new to this field, thanks for your patience and response.

[cz3716087]

Dear author @cz3716087,
Thanks for your wonderful work.
I just want to confirm if there is no code for training the deep research agentic system.
Are the open_data.jsonl you mentioned the training dataset for training the agentic system? And are the rewritten_result.jsonl you mentioned the corresponding training dataset we can generate by the Agentic Dataset Construction module? If we follow the same settings with you, the rewritten_result.jsonl should be the same as open_data.jsonl, am I right?
Hope to get the response from your side. Thanks!

Hi @JadeCityYC,
Thank you for your interest in the project.
Regarding the training code for the deep research agentic system, we haven't provided dedicated scripts in the repository. For Supervised Fine-Tuning (SFT), you can use frameworks like Megatron or LLaMA-Factory and proceed as you would with training other LLM models. For Reinforcement Learning (RL), it's a bit more involved, as you'll need to set up the tool environment yourself (which isn't included in our system). You might want to explore libraries like VERL to build it from scratch.
As for the datasets: open_data.jsonl is generated from the KnowledgeGraphConstruction module and serves as the input for trajectory generation. You'll then need to use your preferred model with the TrajectoryGenerationPipeline to create the trajectories, which will produce rewritten_result.jsonl as the output. Therefore, open_data.jsonl is the input data used to generate rewritten_result.jsonl.
If you have any further questions, feel free to ask!

Dear author @cz3716087,

Thanks for your quick response.

Based on your response, so rewritten_result.jsonl is the training dataset for the deep research agentic system, am I right?

Because you have released the pre-trained 32B model in HuggingFace, does that mean I can load the model and fine-tune the parameters of the agentic system using your existing code?

I am new to this field, thanks for your patience and response.

Yep, the rewritten_result.jsonl is the dataset for the agent model in the deep research agentic system. If you've already generated a new rewritten_result.jsonl, you can use it as training data for further post-training on our 32B model (or any other model you prefer, like a 7B model).

There are some nice papers you might be interested in for understanding foundational concepts in agentic systems (maybe you've already read them):

[https://arxiv.org/abs/2210.03629](ReAct: Synergizing Reasoning and Acting in Language Models - Introduces interleaving reasoning and tool actions to improve task performance)

[https://arxiv.org/abs/2503.16416](Survey on Evaluation of LLM-based Agents - Discusses evaluation methods for agents, including tool-calling capabilities and benchmarks)

[https://arxiv.org/abs/2405.17935](Tool Learning with Large Language Models: A Survey - Comprehensive review of tool learning paradigms, challenges, and future directions)

[JadeCityYC]

Dear author @cz3716087,
Thanks for your wonderful work.
I just want to confirm if there is no code for training the deep research agentic system.
Are the open_data.jsonl you mentioned the training dataset for training the agentic system? And are the rewritten_result.jsonl you mentioned the corresponding training dataset we can generate by the Agentic Dataset Construction module? If we follow the same settings with you, the rewritten_result.jsonl should be the same as open_data.jsonl, am I right?
Hope to get the response from your side. Thanks!

Hi @JadeCityYC,
Thank you for your interest in the project.
Regarding the training code for the deep research agentic system, we haven't provided dedicated scripts in the repository. For Supervised Fine-Tuning (SFT), you can use frameworks like Megatron or LLaMA-Factory and proceed as you would with training other LLM models. For Reinforcement Learning (RL), it's a bit more involved, as you'll need to set up the tool environment yourself (which isn't included in our system). You might want to explore libraries like VERL to build it from scratch.
As for the datasets: open_data.jsonl is generated from the KnowledgeGraphConstruction module and serves as the input for trajectory generation. You'll then need to use your preferred model with the TrajectoryGenerationPipeline to create the trajectories, which will produce rewritten_result.jsonl as the output. Therefore, open_data.jsonl is the input data used to generate rewritten_result.jsonl.
If you have any further questions, feel free to ask!

Dear author @cz3716087,
Thanks for your quick response.
Based on your response, so rewritten_result.jsonl is the training dataset for the deep research agentic system, am I right?
Because you have released the pre-trained 32B model in HuggingFace, does that mean I can load the model and fine-tune the parameters of the agentic system using your existing code?
I am new to this field, thanks for your patience and response.

Yep, the rewritten_result.jsonl is the dataset for the agent model in the deep research agentic system. If you've already generated a new rewritten_result.jsonl, you can use it as training data for further post-training on our 32B model (or any other model you prefer, like a 7B model).

There are some nice papers you might be interested in for understanding foundational concepts in agentic systems (maybe you've already read them):

[https://arxiv.org/abs/2210.03629](ReAct: Synergizing Reasoning and Acting in Language Models - Introduces interleaving reasoning and tool actions to improve task performance)

[https://arxiv.org/abs/2503.16416](Survey on Evaluation of LLM-based Agents - Discusses evaluation methods for agents, including tool-calling capabilities and benchmarks)

[https://arxiv.org/abs/2405.17935](Tool Learning with Large Language Models: A Survey - Comprehensive review of tool learning paradigms, challenges, and future directions)

Dear author @cz3716087,

Thank you so much for your detailed explanations and paper recommendations.
I will follow what you advised accordingly. Thanks!

[JadeCityYC]

Dear author,

Thanks for your wonderful work again. I still have one more question regarding the paper content linking with the code module.

For the KnowledgeGraphConstruction module, I believe it is related to Sections 3.1.1 and 3.1.2 in the paper. For the TrajectoryGenerationPipeline module, is it related to Section 3.1.3, correct? For Section 3.2 in the paper, I could not locate where the related code is. Could you help clarify that for me? Thanks!