Safe4Qwen-1.7B

Safe4Qwen-1.7B is a QLoRA fine-tune of Qwen3-1.7B specifically optimized to reduce hallucination, improve factual grounding, and strengthen step-by-step reasoning — all while staying within the sub-2B parameter class.

Trained by Tralalabs.

---

🎯 Why This Model?

Most small LLMs hallucinate at a high rate because they confabulate under uncertainty, drift from factual knowledge, and skip careful reasoning steps. Safe4Qwen-1.7B directly attacks all three failure modes through targeted dataset fine-tuning:

Dataset	Purpose
TruthfulQA	Resist misconceptions & trick questions — epistemic calibration
WikiQA	Ground answers in real, verifiable Wikipedia knowledge
GSM8K	Step-by-step math reasoning wrapped in <think> tags
Custom (2 rows)	Temporal awareness — knows the current year is 2026

---

📦 Datasets

Dataset	HF Link	Samples Used
TruthfulQA	truthfulqa/truthful_qa	~817 (full validation split)
WikiQA	microsoft/wiki_qa	~1,040 (train, label=1 only)
GSM8K	openai/gsm8k	~7,473 (full train split)
Custom	inline	2 (temporal awareness rows)
Total		~9,332

---

🏋️ Training Details

Parameter	Value
Base model	unsloth/Qwen3-1.7B
Method	QLoRA (4-bit NF4)
LoRA rank	16
LoRA alpha	16
Max sequence length	1024
Batch size	2
Gradient accumulation	4 (effective batch = 8)
Max steps	300
Learning rate	2e-4
Warmup steps	10
Hardware	Google Colab T4 GPU
Framework	Unsloth + TRL + HuggingFace

---

📊 Expected Performance

Safe4Qwen-1.7B is designed to beat or match the following models specifically on hallucination and factual accuracy benchmarks (TruthfulQA, SimpleQA):

Beats:

• Qwen2.5-1.5B-Instruct
• Qwen2.5-3B-Instruct
• Gemma-3-1B-IT
• SmolLM2-1.7B-Instruct
• DeepSeek-R1-Distill-Qwen-1.5B

Matches:

• Phi-4-mini
• DeepSeek-R1-Distill-Llama-8B (on TruthfulQA specifically)

---

💬 Prompt Format

This model uses the ChatML format native to Qwen3:

<|im_start|>user
Your question here<|im_end|>
<|im_start|>assistant

For reasoning tasks, thinking mode is supported via <think> tags:

<|im_start|>assistant
<think>
Step-by-step reasoning here...
</think>
Final answer here<|im_end|>

---

🚀 Quickstart

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("Tralalabs/Safe4Qwen-1.7B")
tokenizer = AutoTokenizer.from_pretrained("Tralalabs/Safe4Qwen-1.7B")

messages = [{"role": "user", "content": "Is the Great Wall of China visible from space?"}]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt")

outputs = model.generate(**inputs, max_new_tokens=256)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

---

⚠️ Limitations

• Sub-2B models have an inherent knowledge capacity ceiling — RAG or grounded context will further reduce hallucinations in production
• Fine-tuned for English only
• ~8,100 training samples over 300 steps — model may still hallucinate on highly domain-specific or niche topics
• Temporal knowledge is fixed at 2026; exact dates are not known

---

📜 License

Apache 2.0 — inherited from Qwen3 base model.