qwen-to-gemma-math — GSM8K Math Reasoning by NEO

Autonomously designed, trained, and evaluated by NEO — Your AI Agent. Try NEO →

---

NEO vs Baseline

Model	GSM8K Accuracy	Correct / 200
google/gemma-4-E2B-it (baseline)	71.0%	142 / 200
daksh-neo/qwen-to-gemma-math (NEO)	75.0%	150 / 200

NEO's model outperforms the baseline by +4.0 percentage points on GSM8K — trained autonomously from scratch using knowledge distillation from Qwen3-plus and LoRA fine-tuning on the full GSM8K training set.

---

Training Pipeline

---

Overview

This model distills mathematical reasoning from Qwen3-plus (teacher) into Gemma 4 2B (student) using chain-of-thought behavioral cloning, followed by LoRA supervised fine-tuning on GSM8K.

NEO autonomously:

1. Designed the full distillation + fine-tuning pipeline
2. Generated 500 chain-of-thought traces from Qwen3-plus via OpenRouter
3. Fine-tuned Gemma 4 2B with LoRA (r=16, alpha=32) on 7,473 GSM8K samples
4. Evaluated and benchmarked against the official baseline
5. Achieved 75% GSM8K accuracy — surpassing the baseline by +4.0%

Architecture: Gemma4ForConditionalGeneration (2B parameters, bfloat16)

---

Distillation Pipeline

Stage 1 — CoT Trace Generation

• Teacher: Qwen3-plus via OpenRouter API (temperature=0, deterministic)
• Prompt: System prompt enforcing step-by-step arithmetic + 2-shot GSM8K examples
• Scale: 500 GSM8K training problems
• Filtering: Length > 25 words + numeric answer + < 30% repeated lines
• Yield: 96.5% valid traces

Stage 2 — LoRA Fine-Tuning

• Method: LoRA r=16, alpha=32
• Dataset: Full GSM8K train split (7,473 samples)
• Epochs: 3 · effective batch size 16 · LR 2e-4 cosine
• Result: 75.0% GSM8K (+4.0% over baseline)

---

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model_id = "daksh-neo/qwen-to-gemma-math"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto"
)
model.eval()

prompt = """Problem: Janet has 3 apples. She gives 1 to her friend and buys 5 more. How many does she have?

Solve step-by-step. End with "The answer is: <number>".

Solution:"""

inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
with torch.no_grad():
    out = model.generate(**inputs, max_new_tokens=256, do_sample=False)
print(tokenizer.decode(out[0], skip_special_tokens=True))

Note: Requires transformers ≥ 5.6.0.dev0 (from source) for Gemma4ForConditionalGeneration support.

---

Evaluation Details

Parameter	Value
Dataset	GSM8K test set
Samples	200
Batch size	16
Precision	bfloat16
Decoding	Greedy (do_sample=False)
Matching	Strict numeric match
GPU	Tesla V100-SXM2-16GB

---

How It Was Built

NEO autonomously designed and executed the full pipeline — zero manual intervention.

1. Identified knowledge distillation as the optimal strategy for math reasoning transfer
2. Prompted Qwen3-plus at temperature=0 to generate deterministic CoT traces
3. Filtered traces for quality (length, answer presence, repetition)
4. Fine-tuned with LoRA on the full GSM8K training set
5. Benchmarked against the official baseline → +4.0% improvement

Try NEO →