🧠 Qwen3-4B-AgentCoder

A Fine-Tuned Model for Enhanced Tool Calling, Code Generation, and Reasoning

• Author: Bruno Pistone
• Repository: br1-pist/Qwen3-4B-AgentCoder
• License: Apache 2.0
• Base model: Qwen/Qwen3-4B-Thinking-2507

---

Model Description

Qwen3-4B-AgentCoder is a fine-tuned version of the Qwen3-4B-Thinking-2507 model, optimized for:

• 🧮 Complex reasoning tasks
• 🧰 Tool calling
• 💻 Code generation

The model was developed through sequential fine-tuning, followed by a Direct Preference Optimization (DPO) post-training stage to improve alignment, coherence, and reasoning accuracy.

Highlights

• Fine-tuned on three specialized datasets
• Retains thinking-mode behavior with long-context reasoning (~264K tokens)
• Post-trained with DPO using chosen/rejected pairs for better alignment
• Excellent balance between tool use, code generation, and reasoning

---

🚀 Direct Use

Qwen3-4B-AgentCoder can be used directly for:

• ✅ Tool calling in complex reasoning tasks
• ✅ Code generation for Python, JS, and other languages
• ✅ Multi-domain reasoning (math, logic, Q&A)

⚠️ Out-of-Scope Use

• ❌ Highly sensitive or confidential data
• ❌ Domains requiring expert-level specialization
• ❌ Tasks where full explainability is mandatory

---

💻 Getting Started

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_name = "br1-pist/Qwen3-4B-AgentCoder"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
model_name,
torch_dtype="auto",
device_map="auto"
)

prompt = "Give me a short introduction to large language models."
messages = [{"role": "user", "content": prompt}]

text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
model_inputs = tokenizer([text], return_tensors="pt").to(model.device)

generated_ids = model.generate(**model_inputs, max_new_tokens=1024)
output = tokenizer.decode(generated_ids[0], skip_special_tokens=True)

print(output)

---

🧠 Training Details

Training Procedure

Phase 1 — Supervised Fine-Tuning

• Learning rate: 1e-5
• Batch size: 4
• Gradient accumulation: 4
• Epochs: 3
• Warmup steps: 100
• Weight decay: 0.01
• Sequence length: ~2.4K tokens

Training Data

• interstellarninja/hermes_reasoning_tool_use (~51K) — multi-turn tool use

Phase 2 — Sequential Fine-Tuning (Supervised)

• Learning rate: 3e-5
• Batch size: 1
• Gradient accumulation: 8
• Epochs: 2
• Warmup steps: 100
• Weight decay: 0.05
• Sequence length: ~13K tokens

Training Data

• ise-uiuc/Magicoder-OSS-Instruct-75K (~38K) — code generation
• open-thoughts/OpenThoughts-114k (~37K) — general reasoning
• interstellarninja/hermes_reasoning_tool_use (~30K) — tool use
• custom/dpo-toolcode-alignment (~15K) — DPO preference pairs

Phase 3 — Post-Training - Direct Preference Optimization (DPO)

After sequential fine-tuning, the model underwent a DPO phase to enhance response alignment, reasoning robustness, and factual consistency.

• Learning rate: 1e-6
• Batch size: 2
• Gradient accumulation: 8
• Epochs: 5
• Beta: 0.2
• Loss type: sigmoid
• Warmup steps: 2
• Sequence length: ~1.5K tokens

DPO Data

• ~1.5K chosen/rejected response pairs
• Rejected samples synthetically generated to represent poor or incoherent answers
• Chosen samples verified or automatically ranked for quality and correctness

Objective

• Encourage the model to prefer chosen completions
• Improve clarity, correctness, and helpfulness
• Reduce hallucinations and verbosity

---

📊 Evaluation

The model was evaluated on multiple benchmarks to assess its capabilities across different domains:

Benchmark	Score	Details
HumanEval	72.0%	Base tests
HumanEval+	68.5%	Base + extra tests
GSM8K	82.0%	1082/1319 correct
MMLU	77.7%	1190/1531 correct (validation split)
Multi-turn Tool Calling	70.0%	70/100 correct

Evaluation Datasets

• HumanEval/HumanEval+: openai/openai_humaneval - 164 hand-written programming problems
• GSM8K: openai/gsm8k (test split) - 1,319 grade school math word problems
• MMLU: cais/mmlu (validation split) - 1,531 multiple-choice questions across 57 subjects
• Tool Calling: Custom dataset - 100 tool calling scenarios

Evaluation Factors

• Tool calling accuracy
• Code generation quality
• General reasoning performance
• Alignment and factual consistency (post-DPO)

Observations

• DPO improved reasoning precision and response coherence
• Code generation accuracy increased in structured programming tasks
• Reduced non-determinism in multi-step tool use

---

🖥️ Technical Specifications

Model Architecture

• Model type: Causal language model
• Parameters: 4.0B
• Context length: ~264K tokens
• Thinking mode: Enabled

Compute Infrastructure

Hardware

• GPU: NVIDIA H100 (80 GB VRAM)
• System RAM: 2 TiB
• Memory per vCPU: 10.67 GiB

Software

• Python: 3.12
• Transformers: 4.55.0
• Libraries: bitsandbytes, safetensors, torch, trl, scikit-learn, tokenizers, psutil, py7zr

---

🧾 Citation

BibTeX

@article{qwen3-4b-thinking-2507-toolcode,
  title={Qwen3-4B-AgentCoder: A Fine-Tuned Model for Enhanced Tool Calling, Code Generation, and Reasoning},
  author={Bruno Pistone},
  year={2025},
  journal={Hugging Face Model Hub}
}

APA

Bruno Pistone. (2025). Qwen3-4B-AgentCoder: A Fine-Tuned Model for Enhanced Tool Calling, Code Generation, and Reasoning. Hugging Face Model Hub. https://huggingface.co/br1-pist/Qwen3-4B-AgentCoder

---

🧭 Recommendations

• Tool use accuracy depends on task complexity
• Code generation may occasionally produce minor syntax issues
• Reasoning strongest in structured, logical, and mathematical contexts
• Avoid using this model for confidential or safety-critical applications

---

🧠 Qwen3-4B-AgentCoder — created by Bruno Pistone
Enhanced reasoning, tool calling, and code generation — refined with DPO alignment