Koishi-1.5

Koishi 1.5

Koishi 1.5 is an updated version of our Koishi model, fine-tuned specifically to augment conversational data by generating Chain of Thought (CoT) reasoning. It is built upon Qwen 2.5 3B Instruct.

Given an input/output pair, Koishi generates a CoT trace.

Use Cases

• Updating older datasets with reasoning traces.
• Adding Chain of Thought to instruct model responses for training reasoning models.
• Generating CoT for model responses where the true reasoning process is unavailable.

Chat Template

The model expects the following structure. Note that Koishi is trained to always begin its generation with Sure, here's the chain of thought:.

Example:

<|im_start|>system
Generate a Chain of Thought chain.<|im_end|>
<|im_start|>user
Input: Where is Paris?
Response: France<|im_end|>
<|im_start|>assistant

Example Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "LucidityAI/Koishi-1.5"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto")

input_text = "What is the capital of France?"
response_text = "Paris"

messages = [
    {"role": "system", "content": "Generate a Chain of Thought chain."},
    {"role": "user", "content": f"Input: Where is Paris?\nResponse: France"}
]

inputs = tokenizer.apply_chat_template(messages, return_tensors="pt", add_generation_prompt=True).to(model.device)

outputs = model.generate(inputs, max_new_tokens=256, do_sample=True)
print(tokenizer.decode(outputs[0][inputs.shape[-1]:], skip_special_tokens=True))