Model Description

This model is fine-tuned to solve Countdown problems: given a target number and a set of allowed numbers, produce a valid arithmetic equation that equals the target. The model generates chain-of-thought reasoning inside <think> tags and a final answer inside <answer> tags.

Training Details

Loss is computed only on response tokens (assistant turn), with prompt tokens masked out.

Evaluation

Evaluated on asingh15/countdown_tasks_3to4 test split (40 prompts) using vLLM with temperature 0.6, top-k 20, top-p 0.95, sampling K=16 responses per prompt.

Usage

python
Copy code
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("ba144220/cs224r-default-project-sft")
tokenizer = AutoTokenizer.from_pretrained("ba144220/cs224r-default-project-sft")

messages = [{"role": "user", "content": "Using the numbers [44, 19, 35], create an equation that equals 98."}]
input_text = tokenizer.apply_chat_template(messages, add_generation_prompt=True, tokenize=False)
inputs = tokenizer(input_text, return_tensors="pt").to(model.device)

outputs = model.generate(**inputs, max_new_tokens=1024, temperature=0.6, top_k=20, top_p=0.95, do_sample=True)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))

Example Output

Prompt: Using the numbers [44, 19, 35], create an equation that equals 98.

Model output:

markdown
Copy code
<think>
Let me try to find a path to 98.
First, let's look at the numbers: 44, 19, and 35.
Working backwards from 98: 98 is even, so it might be helpful to look for factors.
Let's try some operations:
44 + 19 = 63 (getting closer)
63 + 35 = 98 (this is our target!)
We can do: 19 + 44 = 63, 63 + 35 = 98.
Let's verify: (19 + 44) + 35 = 98 ✓
</think>
<answer> (19 + 44) + 35 </answer>

Score: 1.0 (correct)

Limitations

• Trained and evaluated only on the Countdown arithmetic task; not intended for general-purpose use.
• Performance degrades on harder problems with more numbers or larger targets.
• The 0.5B parameter size limits reasoning capacity compared to larger models.

Authors

Yuchi Hsu (yuchihsu@stanford.edu) and Ryan He (ryanhe@stanford.edu), Stanford CS224R Spring 2026.