I've been experimenting with a new quantization approach that selectively elevates the precision of key layers beyond what the default IMatrix configuration provides.

In my testing, standard IMatrix quantization underperforms at lower bit depths, especially with Mixture of Experts (MoE) models. To address this, I'm using the --tensor-type option in llama.cpp to manually "bump" important layers to higher precision. You can see the implementation here:
👉 Layer bumping with llama.cpp

While this does increase model file size, it significantly improves precision for a given quantization level.

I'd love your feedback—have you tried this? How does it perform for you?

Click here to get info on choosing the right GGUF model format

Model Card for PokeeResearch

Model Details

Model Description

PokeeResearch-7B is a 7-billion-parameter deep research agent developed by Pokee AI to advance reliable, aligned, and scalable research-grade reasoning in tool-augmented LLMs.
The model integrates Reinforcement Learning from AI Feedback (RLAIF) with a robust reasoning scaffold, enabling it to conduct complex, multi-step research workflows that include self-correction, verification, and synthesis across multiple independent research threads.

Developed by: Pokee AI
Model type: Tool-augmented large language model (LLM) research agent
Language(s): English, Chinese and many more
License: Apache 2.0
Finetuned from model: Qwen2.5-7B-Instruct

Model Sources

Repository: https://github.com/Pokee-AI/PokeeResearchOSS
Paper: PokeeResearch: Effective Deep Research via Reinforcement Learning from AI Feedback and Robust Reasoning Scaffold, Pokee AI, October 2025
Project Page: https://pokee.ai/deepresearch-preview

Uses

Direct Use

PokeeResearch-7B is designed for deep research automation, where the model autonomously:

Decomposes complex user queries
Retrieves and reads from external sources
Synthesizes factual, verifiable, and grounded answers

It can be used as a standalone research assistant or integrated into multi-agent systems to support academic, enterprise, or product-level research tasks.

Downstream Use

PokeeResearch-7B can be fine-tuned or extended for:

Domain-specific scientific discovery
Autonomous document retrieval and synthesis
Multi-source verification and summarization pipelines
Integration into reinforcement learning research agents (RLHF/RLAIF frameworks)

Out-of-Scope Use

The model should not be used for:

Generating unverified or speculative claims
Automated decision-making in high-stakes domains (medical, legal, or financial)
Applications requiring strict factual precision without external verification
Generating content without citation or evidence tracing

Bias, Risks, and Limitations

PokeeResearch-7B is optimized for factual grounding and robustness, but limitations include:

Dependence on external data quality and retrieval accuracy
Potential semantic bias introduced by AI-based feedback signals
Limited coverage for non-English or multi-modal reasoning tasks
Risk of hallucinated synthesis when sources conflict or lack clarity

Recommendations

Users should:

Cross-verify answers, especially in multi-hop reasoning cases
Monitor output for citation accuracy and alignment with source data
Refrain from using outputs as sole evidence in decision-critical contexts

How to Get Started with the Model

please refer to the following codebase for how to use PokeeResearch-7B https://github.com/Pokee-AI/PokeeResearchOSS/blob/main/README.md

Training Details

Training Data

Dataset: MiroRL-GenQA dataset (MiroMind AI, 2025)
Data characteristics: Complex, multi-turn question–answer pairs requiring multi-step reasoning
Data filtering: No benchmark data used for testing; the model was trained only on open-domain text Q&A samples

Training Procedure

Preprocessing

Normalization and tokenization aligned with Qwen2.5 tokenizer
Structured prompt–response pairs in research/verification format (<tool_call>, <answer>, <verification>)

Training Hyperparameters

Algorithm: RLOO (REINFORCE Leave-One-Out)
Batch size: 64
Research threads per prompt: 8
Learning rate: 3e-6
Context limit: 32,768 tokens
Steps: 140 fine-tuning iterations
Regularization: None (no entropy or KL regularization)
Precision regime: bf16 mixed precision

Reward Design

Combined reward signal from:
- AI feedback (semantic equivalence via external LLM judge)
- Format adherence reward (ensures correct agent behavior)

Speeds, Sizes, Times

Model size: 7 billion parameters
Training duration: ~5 days on 8 × A100 80G GPUs
Checkpoint size: ~13 GB

Evaluation

Testing Data, Factors & Metrics

Testing Data

10 open-domain research and QA benchmarks:

NQ, TriviaQA, PopQA, HotpotQA, 2WikiMultiHopQA, Musique, Bamboogle, GAIA, BrowseComp, Humanity’s Last Exam

Factors

Benchmarks differ by reasoning depth, retrieval dependence, and factual precision requirements.
Evaluations disaggregate by dataset difficulty and task type (single-hop vs multi-hop).

Metrics

Mean accuracy (mean@4 across independent research threads) based on

Results

PokeeResearch-7B (RTS variant) and PokeeResearch-7B outperforms all baselines at 7B scale across 10 benchmarks.
Highlights (mean@4 accuracy):

Method	HLE	GAIA	BrowseComp	BAMB	2WIKI	TQ	NQ	POPQA	MUSIQUE	HOTPOTQA
R1searcher	5.4	8.3	1.0	63.2	61.4	77.2	59.6	51.8	35.8	62.4
SearchR1	13.0	18.7	0.4	67.8	62.8	81.0	67.6	59.6	33.2	63.2
ZeroSearch	8.6	9.9	1.4	51.4	33.6	61.6	48.2	38.0	19.0	32.4
ASearcher	13.8	22.1	3.2	68.8	69.2	85.2	71.2	58.2	35.8	71.0
DeepResearcher	6.0	24.03	1.8	71.0	58.8	82.2	60.2	55.2	26.8	56.6
PR	15.2	36.9	5.4	74.5	74.0	91.3	75.1	59.8	39.8	71.2
PR+	17.6	41.3	8.4	75.0	75.0	91.8	75.0	60.0	41.4	71.6

Summary

PokeeResearch-7B variants achieves state-of-the-art performance among 7B-scale open deep research agents, validating RLAIF and reasoning scaffold design for robust, verifiable research workflows.

Technical Specifications

Model Architecture and Objective

Base Architecture: Transformer decoder (Qwen2.5-7B-Instruct backbone)
Objective: Reinforcement learning with AI feedback to maximize semantic correctness and alignment with human-style reasoning

Compute Infrastructure

Hardware

NVIDIA A100 80GB GPUs ×8 for training and x1 for inference

Citation

BibTeX:

bibtex

@article{pokee2025deepresearch,
  title={PokeeResearch: Effective Deep Research via
          Reinforcement Learning from AI Feedback and Robust Reasoning Scaffold},
  author={Yi Wan* and Jiuqi Wang* and Liam Li
          and Jinsong Liu and Ruihao Zhu and Zheqing Zhu},
  journal={Pokee AI Technical Report},
  year={2025},
  url={https://arxiv.org/pdf/2510.15862}
}

APA: Wan, Y., Wang, J., Li, L., Liu, J., Zhu, R., & Zhu, Z. (2025). PokeeResearch: Effective Deep Research via Reinforcement Learning from AI Feedback and Robust Reasoning Scaffold. Pokee AI.

Glossary

RLAIF: Reinforcement Learning from AI Feedback – optimization using LLM-based reward signals.
RLOO: REINFORCE Leave-One-Out – unbiased policy gradient variant for on-policy learning.
RTS: Research Threads Synthesis – synthesis of multiple independent reasoning threads at inference time.

More Information

For technical details, visit: https://github.com/Pokee-AI/PokeeResearchOSS
For inquiries, contact: hello@pokee.ai

Model Card Authors

Yi Wan, Jiuqi Wang, Liam Li, Jinsong Liu, Ruihao Zhu, and Zheqing Zhu — Pokee AI Research Team

Model Card Contact

Pokee AI Team — hello@pokee.ai

🚀 If you find these models useful

Help me test my AI-Powered Quantum Network Monitor Assistant with quantum-ready security checks:

👉 Quantum Network Monitor

The full Open Source Code for the Quantum Network Monitor Service available at my github repos ( repos with NetworkMonitor in the name) : Source Code Quantum Network Monitor. You will also find the code I use to quantize the models if you want to do it yourself GGUFModelBuilder

💬 How to test:
Choose an AI assistant type:

TurboLLM (GPT-4.1-mini)
HugLLM (Hugginface Open-source models)
TestLLM (Experimental CPU-only)

What I’m Testing

I’m pushing the limits of small open-source models for AI network monitoring, specifically:

Function calling against live network services
How small can a model go while still handling:
- Automated Nmap security scans
- Quantum-readiness checks
- Network Monitoring tasks

🟡 TestLLM – Current experimental model (llama.cpp on 2 CPU threads on huggingface docker space):

✅ Zero-configuration setup
⏳ 30s load time (slow inference but no API costs) . No token limited as the cost is low.
🔧 Help wanted! If you’re into edge-device AI, let’s collaborate!

Other Assistants

🟢 TurboLLM – Uses gpt-4.1-mini :

**It performs very well but unfortunatly OpenAI charges per token. For this reason tokens usage is limited.
Create custom cmd processors to run .net code on Quantum Network Monitor Agents
Real-time network diagnostics and monitoring
Security Audits
Penetration testing (Nmap/Metasploit)

🔵 HugLLM – Latest Open-source models:

🌐 Runs on Hugging Face Inference API. Performs pretty well using the lastest models hosted on Novita.

💡 Example commands you could test:

"Give me info on my websites SSL certificate"
"Check if my server is using quantum safe encyption for communication"
"Run a comprehensive security audit on my server"
'"Create a cmd processor to .. (what ever you want)" Note you need to install a Quantum Network Monitor Agent to run the .net code on. This is a very flexible and powerful feature. Use with caution!

Final Word

I fund the servers used to create these model files, run the Quantum Network Monitor service, and pay for inference from Novita and OpenAI—all out of my own pocket. All the code behind the model creation and the Quantum Network Monitor project is open source. Feel free to use whatever you find helpful.

If you appreciate the work, please consider buying me a coffee ☕. Your support helps cover service costs and allows me to raise token limits for everyone.

I'm also open to job opportunities or sponsorship.

Thank you! 😊

Mungert/pokee_research_7b-GGUF

pokee_research_7b GGUF Models

Model Generation Details

Quantization Beyond the IMatrix