sinhal/qwen3-4b-instruct-2507-gsm8k-reasoning

Qwen3-4B-Instruct-2507 - LoRA on gsm8k-reasoning

A QLoRA adapter for Qwen/Qwen3-4B-Instruct-2507, trained on thesven/gsm8k-reasoning.

This is an adapter, not a full model. Load it on top of the base.

Training

Rows seen	6,714 (the whole dataset less the holdout; 5,000 were planned)
Sessions	1
Final training loss	0.1456
Method	QLoRA, 4-bit NF4, r=16, alpha=32
Hardware	Kaggle T4 (16GB), free tier

Does it help?

On 200 rows held out of training and never seen by the model, mean token loss fell from 2.3668 to 0.1418 (94.0% lower). The same rows were scored before and after training, so this is a like-for-like comparison rather than a training-loss curve.

Held-out loss, base model	2.3668
Held-out loss, this adapter	0.1418
Loss reduction	2.2250 (94.0% lower)
Held-out rows	200

Trained in 1 session on Kaggle's free GPU, which stops at twelve hours - longer runs resume from the previous session's checkpoint.

Use

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-4B-Instruct-2507")
model = PeftModel.from_pretrained(base, "sinhal/qwen3-4b-instruct-2507-gsm8k-reasoning")
tok = AutoTokenizer.from_pretrained("Qwen/Qwen3-4B-Instruct-2507")

Honest limitations

Trained on one dataset for 1 session on free-tier hardware. It has not been benchmarked, and no claim is made that it beats the base model at anything. It exists because the pipeline that produced it runs unattended, and it is published so the work is inspectable.

Built and published automatically by an agent. The base model and dataset licences above govern use.