Qwen3-1.7B-ft-bf16

Qwen3-1.7B-ft-bf16 is a fine-tuned, moderately abliterated variant of the Qwen3-1.7B model. Built upon the robust Qwen3 architecture, this version emphasizes improved context awareness and moderate behavioral flexibility, while maintaining high standards in reasoning, instruction-following, and multilingual performance. It is designed to perform consistently across general-purpose dialogue, technical reasoning, creative writing, and multilingual tasks.

Key Highlights:

• Improved Context Awareness: Retains and utilizes long-span contextual information effectively, making it suitable for long conversations, document analysis, and summarization.
• Moderate Abliteration: Introduces controlled experimental freedoms for enhanced expressiveness and adaptability, while preserving safety and alignment.
• Dual-Mode Thinking Support: Supports dynamic switching between deep logical reasoning and efficient casual dialogue, making it task-aware and context-adaptive.
• Multilingual Excellence: Robust across 100+ languages, handling translation, multilingual instruction, and language-specific tasks seamlessly.
• Tool and Agent Integration: Performs well in agent-driven scenarios and can interface with tools and APIs in both thinking and non-thinking modes.

---

Quickstart with 🤗 Transformers

pip install transformers==4.51.3
pip install huggingface_hub[hf_xet]

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "prithivMLmods/Qwen3-1.7B-ft-bf16"

# Load tokenizer and model
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype="auto",
    device_map="auto"
)

# Define prompt and apply chat template
prompt = "Explain why the sky appears blue during the day and red at sunset."
messages = [{"role": "user", "content": prompt}]
text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
    enable_thinking=True
)

# Tokenize input
model_inputs = tokenizer([text], return_tensors="pt").to(model.device)

# Generate response
generated_ids = model.generate(
    **model_inputs,
    max_new_tokens=32768
)
output_ids = generated_ids[0][len(model_inputs.input_ids[0]):].tolist()

# Optional: Separate thinking content
try:
    index = len(output_ids) - output_ids[::-1].index(151668)  # token ID for </think>
except ValueError:
    index = 0

thinking_content = tokenizer.decode(output_ids[:index], skip_special_tokens=True).strip("\n")
content = tokenizer.decode(output_ids[index:], skip_special_tokens=True).strip("\n")

print("thinking content:", thinking_content)
print("content:", content)

---

Recommended Settings

• Sampling Parameters:
  • Thinking Mode: temperature=0.6, top_p=0.95, top_k=20, min_p=0.0
  • Non-Thinking Mode: temperature=0.7, top_p=0.8, top_k=20, min_p=0.0
• Max Token Length:
  • Standard Tasks: 32768
  • Complex/Extended Tasks: 38912

---

Prompting Guidelines

• Math Problems:
  "Please reason step by step, and put your final answer within \boxed{}."
• MCQs:
  Format: {"answer": "C"}
• Dialogues:
  Include only final responses in history; omit internal thinking logs for efficiency.