desidero32/Qwen3.5-2B-OpenMathReasoning-Finetuned

Qwen3.5-2B SFT Merged

This model is a merged supervised fine-tuned version of Qwen/Qwen3.5-2B.

It was fine-tuned with QLoRA / LoRA and then merged into the base model for direct inference.

Model Summary

• Base model: Qwen/Qwen3.5-2B
• Fine-tuning method: QLoRA 4-bit + LoRA
• Output type: merged full model
• Primary focus: math reasoning and general reasoning

Training Data

This model was trained on a mixed supervised fine-tuning dataset with emphasis on math reasoning.

Main datasets used include:

• nvidia/OpenMathReasoning
• jasonrqh/Math-CoT-20k

Additional instruction and domain datasets were also used in the training pipeline in this project.

Intended Use

This model is intended for:

• math problem solving
• reasoning-heavy instruction following
• general text generation

It is not specifically optimized for factual QA, safety-critical domains, or tool use.

Benchmark Snapshot

Raw vs merged comparisons from this project:

Benchmark	Raw Base	Merged Model
GSM8K	0.66	0.74
MATH-500	0.27	0.33
ARC-Challenge	0.21	0.29
CommonsenseQA	0.21	0.28
BoolQ	0.75	0.74
WinoGrande	0.52	0.51

These are small project-side comparisons, not an official leaderboard submission.

How to Use

Transformers

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "YOUR_USERNAME/YOUR_MODEL_REPO"

tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16 if torch.cuda.is_available() else torch.float16,
    device_map="auto",
    trust_remote_code=True,
)

messages = [
    {"role": "system", "content": "You are a careful reasoning assistant."},
    {"role": "user", "content": "Solve: If 3x + 5 = 20, what is x? End with Final answer: <answer>"},
]

prompt = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
)

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

with torch.inference_mode():
    outputs = model.generate(
        **inputs,
        max_new_tokens=256,
        do_sample=False,
        temperature=0.0,
        pad_token_id=tokenizer.pad_token_id,
        eos_token_id=tokenizer.eos_token_id,
    )

generated = outputs[0][inputs["input_ids"].shape[1]:]
print(tokenizer.decode(generated, skip_special_tokens=True))

## Limitations

- Performance is strongest on reasoning-style prompts close to the SFT data distribution.
- Gains are not uniform across all reasoning benchmarks.
- Some benchmark improvements may reflect output-format adaptation as well as reasoning improvement.

## Training / Eval Project

This model was trained and evaluated in a local project with raw-vs-merged comparisons on math and reasoning benchmarks including:

- GSM8K
- MATH-500
- Math-CoT-20k
- MMLU math
- BoolQ
- WinoGrande
- CommonsenseQA
- ARC-Challenge