You can deploy this model and start earning money today!

sabia0080/
qwen3-4b-structevalt-lora-hard4k

Name: qwen3-4b-structevalt-lora-hard4k
Availability: InStock

safetensors

Views

No views yet

base_model: "unsloth/Qwen3-4B-Instruct-2507"

dataset: "daichira/structured-hard-sft-4k"

language: "en"

license: "Apache-2.0"

library_name: "peft"

pipeline_tag: "text-generation"

tags: ["qlora","lora","peft","unsloth","structured-output","structeval-t","json","yaml","xml","toml","csv"]

qwen3-4b-structevalt-lora-hard4k

This repository provides a LoRA adapter fine-tuned from unsloth/Qwen3-4B-Instruct-2507 using QLoRA (4-bit, Unsloth).

✅ Adapter-only repository: this repo contains LoRA adapter weights (and tokenizer files if uploaded). You must load the base model separately.

Goal

Improve structured output reliability for StructEval-T style tasks, especially on harder conversion/extraction cases across JSON/XML/YAML/TOML.

During training, loss is applied only to the final assistant output, masking intermediate reasoning (CoT) using output markers (e.g., Output:, Final:).

Training Configuration

Base model: unsloth/Qwen3-4B-Instruct-2507
Dataset: daichira/structured-hard-sft-4k (CC-BY-4.0)
Method: QLoRA (4-bit) + PEFT LoRA
Max sequence length: 512
Epochs: 1
Max steps: 200
Learning rate: 1e-5
Warmup ratio: 0.1
Weight decay: 0.05
Batch size (per device): train 2 / eval 2
Gradient accumulation: 8 (effective batch = 16)
LoRA: r=64, alpha=128, dropout=0.0
LoRA target modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
Seed: 3407
Output learn mode: after_marker
Output markers: Output:, OUTPUT:, Final:, Answer:, Result:, Response:

Run info (for reproducibility)

RUN_ID: 20260205_111047__p1_hard4k__base=Qwen3-4B-Instruct-2507__ds=structured-hard-sft-4k__lr=1e-5__len=512__steps=200__r=64

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
import torch

base_id = "unsloth/Qwen3-4B-Instruct-2507"
adapter_id = "sabia0080/qwen3-4b-structevalt-lora-hard4k"

tokenizer = AutoTokenizer.from_pretrained(base_id)
model = AutoModelForCausalLM.from_pretrained(
    base_id,
    torch_dtype=torch.float16,
    device_map="auto",
)

model = PeftModel.from_pretrained(model, adapter_id)
model.eval()

License & Terms

Dataset license: CC-BY-4.0 (daichira/structured-hard-sft-4k)
Base model terms: comply with the base model’s original terms of use.
Repo license (Apache-2.0) applies to this adapter repository contents (adapter files / model card), and does not override dataset/base-model terms.