🔥 Update (April 5): I’ve released the complete training notebook, codebase, and a comprehensive PDF guide to help beginners and enthusiasts understand and reproduce this model's fine-tuning process.

❤️ Special thanks to the Unsloth open-source library and @KyleHessling1 for their support.

📚 Resources & Guides

👉 GitHub Repository: Jackrong-llm-finetuning-guide Visit the repo to dive into the codebase and reproduce the results locally or on Colab.

📥 Core Technical Document

🔗 Qwopus3.5-27b Complete Fine-Tuning Guide (PDF)

The Full Pipeline: A step-by-step walkthrough—from downloading the base model and unifying heterogeneous data, to configuring trainer hyperparameters and publishing to Hugging Face.
Beginner Friendly: Includes an introductory guide to getting started with Google Colab and Unsloth.
Feedback welcome! If you spot any areas for improvement, please let me know and I will update it promptly.

A Note: My goal isn't just to detail a workflow, but to demystify LLM training. Beyond the social media hype, fine-tuning isn't an unattainable ritual—often, all you need is a Google account, a standard laptop, and relentless curiosity.

No one starts as an expert, but every expert was once brave enough to begin.

All training and testing for this project were self-funded. If you find this model or guide helpful, a Star ⭐️ on GitHub would be the greatest encouragement. Thank you! 🙏

[!Note] The Claude series model optimizations are named under the Qwopus3.5 series, with the latest version being 🌟Qwopus3.5-v3.

💡 Model Introduction

Qwopus3.5-9B-v3 is a reasoning-enhanced model based on Qwen3.5-9B. Its core objective is to simultaneously improve reasoning stability and correctness while optimizing inference efficiency, ultimately achieving stronger cross-task generalization capabilities—particularly in programming.

By continuing to optimize the fundamental structure of its reasoning process alongside high-quality reasoning distillation and structural alignment, it enables the model to achieve higher accuracy rates through shorter, more stable reasoning paths.

🍎 Qwopus3.5-9B-v3: Humaneval Benchmark Evaluation

Inference for models was conducted under the Unsloth runtime environment using bfloat16 (BF16) precision, which provides a balance of numerical range and memory efficiency well-suited to 9B-scale inference. Answer verification, partial chain-of-thought adjudication, and statistical analysis were cross-validated using GPT-5.4-Pro (Thinking) and Claude Opus 4.6 (Thinking) to ensure accuracy and reproducibility of the evaluation outcomes.

HumanEval
I evaluated three 9B-scale Qwen-family models on the full 164-task HumanEval benchmark under a task-level adjudication protocol that resolves code-extraction pollution, answer/code separation issues, and clearly inferable truncated outputs using raw generations. Under this fair and strict evaluation setting, Qwopus3.5-9B-v3 achieves the best base pass@1 of 87.80% (144/164), outperforming both Qwen3.5-9B (82.93%, 136/164) and Claude-Distilled-v2 (82.32%, 135/164). Furthermore, on the stricter plus pass@1 evaluation, Qwopus3.5-9B-v3 also extends its lead to 82.93% (136/164) compared to 77.44% (127/164) for the official baseline (+5.49 pp) and 78.66% (129/164) for the distilled variant.

Model	Base pass@1	Plus pass@1	Rescues (From GPT)	Improvement vs Qwen3.5-9B
Qwopus3.5-9B-v3	87.80% (144/164)	82.93% (136/164)	1	📈 Base: +4.87 pp / Plus: +5.49 pp
Qwen3.5-9B	82.93% (136/164)	77.44% (127/164)	2	Baseline
Claude-Distilled-v2	82.32% (135/164)	78.66% (129/164)	0	📉 Base: -0.61 pp / 📈 Plus: +1.22 pp vs Qwen3.5-9B

Note: The test results presented here differ from the scores on the 9B-v2 model card because the context length was increased for this evaluation. Consequently, the number of tasks affected by context window truncation has changed for each model, leading to different final scores. Please ensure comparisons are made under the same variable settings.

All post-evaluation standard result files will be uploaded to this repository for transparency and reproducibility. These include:

Jackrong_Qwen3.5-9B-Claude-4.6-Opus-Reasoning-Distilled-v2_humaneval_all_evalonly_eval_results
Jackrong_Qwopus3.5-9B-v3-test1_humaneval_all_evalonly_eval_results
qwen_Qwen3.5-9B_humaneval_all_evalonly_eval_results

⚠️ Note on evaluation artifacts.
The released result files are based on raw model generations, which may contain formatting issues (e.g., Markdown wrappers, answer/code mixing), truncation, or minor token-level corruption.

🏃 Qwopus3.5-9B-v3: MMLU-Pro Benchmark Evaluation

I evaluated on 280 MMLU-Pro questions across the following domains: Biology, Chemistry, Computer Science, Health, Mathematics, Physics, and Other Sciences.

All question IDs are identical across both model runs.

Accuracy

Model	Correct	Total	Accuracy
Qwen3.5-9B	225	280	80.36%
Qwopus3.5-9B-v3	229	280	81.79%

Result:
Qwopus3.5-9B-v3 leads by +1.43 pp

Reasoning Efficiency

Metric	Qwen3.5-9B	Qwopus3.5-9B-v3
Avg think length	7116 chars	5313 chars
Passes / 10k chars	1.26	1.66
Chars / correct pass	7938	6032

Reasoning Efficiency Improvements

−25.3% shorter reasoning
+31.7% higher efficiency
−24.0% lower cost per correct answer

Evaluation Summary

While the overall accuracy margin (+1.43 pp) is modest, Qwopus3.5-9B-v3 fundamentally shifts the accuracy-cost paradigm, achieving its victory while spending significantly less reasoning budget. With a 25.3% reduction in mean think length and 24.0% lower token cost per correct answer, this iteration is highly optimized for latency, token budget, and context pressure.

Furthermore, across the mixed domain profile, Qwopus3.5-9B-v3 uniquely offsets Qwen3.5-9B's slight edge in biology, CS, and math by excelling in physics, chemistry, and significantly lowering its unfinished-output rate. Its final rank benefits as much from raw correctness as from an improved ability to cleanly and reliably complete analytical boundaries.

🗺️ Training Pipeline Overview

text

Base Model (Qwen3.5-9B)
 │
 ▼
Qwen3.5-9B fine-tuned with Unsloth
 │
 ▼
Supervised Fine-Tuning (SFT) + LoRA
(Response-Only Training masked on "<|im_start|>assistant\n<think>")
 │
 ▼
Qwopus3.5-9B-v3

🧠 Example of Learned Reasoning Scaffold

The model includes targeted optimizations addressing Qwen3.5's tendency toward excessive or repetitive reasoning on simple queries. By distilling the structured reasoning habits of top-tier models like Claude Opus, Qwopus3.5-9B-v3 adopts a highly organized, step-by-step cognitive layout.

text

Example：The user is asking about [Topic] and how it differs from [Topic B]. This is a [Task type] question. Let me break this down:

1. What is [Topic A]?
   - [Fact/Mechanism 1]
   - [Fact/Mechanism 2]
2. What is [Topic B]?
   - [Fact/Mechanism 1]
3. Key differences:
   - [Comparison Point 1]
   - [Comparison Point 2]

Let me make sure to be accurate: [...]
Actually, I should double-check: is [Fact] used before [Fact]? Yes, typically...
Let me provide a clear, well-structured answer:

📚 Training Data

The model was fine-tuned on a high-fidelity reasoning dataset, which was meticulously curated from a blend of premium open-source sources on Hugging Face. This dataset is the result of a rigorous mixing and cleaning process, specifically designed to filter out low-quality responses and ensure consistently strong logical performance across diverse analytical domains.

(Rest assured, the entire process is strictly by-the-book and 100% compliant with all terms and open-source licenses!)

⚠️ Limitations & Intended Use

Hallucination Risk: While reasoning is strong, the model remains an autoregressive LLM; external facts provided during the thinking sequence may occasionally contain hallucinations if verifying real-world events.
Intended Scenario: Best suited for offline analytical tasks, coding, math, and heavy logic-dependent prompting where the user needs to transparently follow the AI's internal logic.
This model is a test version intended solely for learning and demonstration purposes, and is for academic research and technical exploration use only.

🙏 Acknowledgements

Significant thanks to the Unsloth AI team for making rapid fine-tuning of large LLM models accessible. Additionally, we acknowledge Qwen internally, and the open-source community developers producing exceptional distilled datasets.

This qwen3_5 model was trained 2x faster with Unsloth and Huggingface's TRL library.

📖 Citation

If you use this model in your research or projects, please cite:

bibtex

@misc{jackrong_qwen35_9b_v3
  title        = {Jackrong/Qwopus3.5-9B-v3},
  author       = {Jackrong},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/Jackrong/Qwopus3.5-9B-v3-GGUF}}
}