Qwen3-8B-ft (GGUF Quantized)
Model Details
Model Description
A fine-tuned version of Qwen3-8B, optimized for improved performance across reasoning, code generation, and general knowledge tasks.
- Developed by: Aname-Tommy
- Model type: Causal Language Model (Decoder-only Transformer)
- Language(s) (NLP): English, Japanese, Chinese
- Finetuned from model: Qwen/Qwen3-8B
Model Sources
- Repository: https://huggingface.co/Aname-Tommy/qwen3-8b-ft
Benchmark Results
| Benchmark | Metric | Score |
|---|---|---|
| LiveCodeBench | Hard / Pass@1 | 23.60% (Like-SoTA) |
| GSM8K | 5-shot / Exact Match | 86.20% |
| MMLU | Average / Accuracy | 72.23% |
| HumanEval | Pass@1 | 64.02% |
| HellaSwag | Acc_norm | 76.21% |
| ARC-Challenge | Acc_norm | 56.23% |
Direct Use
This model can be used for general-purpose text generation, question answering, code generation, mathematical reasoning, and other natural language processing tasks.
Bias, Risks, and Limitations
This model may inherit biases present in the base Qwen3-8B model and the fine-tuning data. Users should exercise caution and verify model outputs before relying on them in critical applications.
Recommendations
Users (both direct and downstream) should be made aware of the risks, biases, and limitations of the model. The model should not be used for generating harmful, misleading, or unethical content.
Evaluation
Metrics
| Benchmark | Metric | Description |
|---|---|---|
| LiveCodeBench | Pass@1 (Hard) | Pass rate on competitive programming problems |
| GSM8K | Exact Match (5-shot) | Accuracy on grade-school math word problems |
| MMLU | Accuracy (Average) | Average accuracy across multi-domain knowledge |
| HumanEval | Pass@1 | Pass rate on Python code generation tasks |
| HellaSwag | Acc_norm | Normalized accuracy on commonsense sentence completion |
| ARC-Challenge | Acc_norm | Normalized accuracy on science reasoning questions |
Results
| Benchmark | Score |
|---|---|
| LiveCodeBench (Hard / Pass@1) | 23.60% |
| GSM8K (5-shot / Exact Match) | 86.20% |
| MMLU (Average / Accuracy) | 72.23% |
| HumanEval (Pass@1) | 64.02% |
| HellaSwag (Acc_norm) | 76.21% |
| ARC-Challenge (Acc_norm) | 56.23% |
Summary
The model demonstrates strong performance in mathematical reasoning (GSM8K: 86.20%) and commonsense reasoning (HellaSwag: 76.21%). Code generation capability is solid (HumanEval: 64.02%), and multi-domain knowledge (MMLU: 72.23%) remains stable. The model also shows competitive results on competitive programming benchmarks (LiveCodeBench: 23.60%) and science reasoning (ARC-Challenge: 56.23%).
Technical Specifications
Model Architecture and Objective
Decoder-only Transformer based on the Qwen3-8B architecture. Trained with a causal language modeling (next-token prediction) objective.
Glossary
- Pass@1: The probability that the first generated sample passes all test cases.
- Acc_norm: Normalized accuracy, accounting for sequence length bias.
- Exact Match: Whether the model's output exactly matches the ground truth answer.
Model Card Authors
Aname-Tommy
Model Card Contact
Aname-Tommy