Qwen2.5-1.5B-Reasoning

A QLoRA fine-tune of unsloth/Qwen2.5-1.5B-Instruct-bnb-4bit, trained on a subset of the orca-math-word-problems-200k dataset, aimed at step-by-step math word problem solving.

Part of an ongoing series of small, task-focused fine-tunes exploring what narrow specialization can (and can't) improve in sub-2B parameter models.

Model Details

• Developed by: agentbyumer
• Base model: unsloth/Qwen2.5-1.5B-Instruct-bnb-4bit
• License: apache-2.0
• Language: English
• Trained with: Unsloth

Training Details

Method	QLoRA
Dataset	orca-math-word-problems-200k (subsampled, ~8,000 rows)
Epochs	3
Context length	4,096
Learning rate	1e-5

Intended Use

Step-by-step math word problem solving — the training data consists of question/answer pairs where answers show worked reasoning (e.g. "Let's call the number x... First we need to... So the result is...") rather than bare final answers.

Evaluation

No formal benchmark was run on this model. Informal side-by-side testing against the base model (same prompts, same sampling) on a handful of hand-written word problems showed:

• On simple, single-step problems, both base and fine-tuned models performed similarly, with the fine-tuned model tending toward the plain-text step-by-step formatting style of the training data (vs. the base model's more LaTeX-heavy formatting).
• On harder multi-step problems involving compound relationships or non-integer results, both models made arithmetic errors, and the fine-tuned model did not show a consistent improvement over the base model — in some cases it introduced additional arithmetic slips the base model didn't make.

This is shared transparently rather than omitted: it reflects a real limitation of this specific training run (a relatively small data subset, low learning rate, single training pass) rather than a claim about what the approach is generally capable of. Contributions, re-training with a larger subset, or evaluation on a formal benchmark (e.g. GSM8K) are welcome.

How to Use

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("agentbyumer/qwen2.5-1.5b-reasoning", device_map="auto")
tokenizer = AutoTokenizer.from_pretrained("agentbyumer/qwen2.5-1.5b-reasoning")

A GGUF-quantized version for local/CPU inference is available at agentbyumer/qwen2.5-1.5b-reasoning-GGUF.

---

This qwen2 model was trained 2x faster with Unsloth