SeanWang0027/qwen3-4b-nemotron-if-off-sft-v1
SeanWang0027/qwen3-4b-nemotron-if-off-sft-v1 is a 4 billion parameter Qwen3-based language model fine-tuned for instruction-following tasks, specifically excluding reasoning-heavy prompts. Developed by SeanWang0027, this model leverages a LoRA adapter merged into the base weights, focusing on concise and format-following responses. It demonstrates improved performance on a MATH-500 proxy evaluation, achieving answers more efficiently and with fewer unparsed outputs compared to its base model.
Loading preview...
Model Overview
This model, SeanWang0027/qwen3-4b-nemotron-if-off-sft-v1, is a supervised fine-tune (SFT) of the Qwen/Qwen3-4B-Base model. It was trained using a LoRA adapter, which has been merged into the base weights, resulting in a standard Transformers checkpoint. The fine-tuning specifically targeted instruction-following capabilities while deliberately excluding reasoning-intensive examples.
Training Details
The model was fine-tuned on a subset of the nvidia/Nemotron-Instruction-Following-Chat-v1 dataset. The training data was carefully selected to include only reasoning=off and capability_target=instruction_following conversations, totaling 12,373 single-turn examples. This exclusion aimed to prevent the model from generating lengthy, chatty, or self-identifying responses often found in the chat portion of the dataset. Training utilized bf16 precision and FlashAttention-2, with a learning rate of 1e-4 over one epoch.
Performance Highlights
Evaluation on a MATH-500 proxy split (200 items) showed significant improvements over the base model:
avg@8score: Increased from 0.3900 (base) to 0.6312 (this model).pass@8score: Improved from 0.720 (base) to 0.785.- Answer parsing: Unparsed answers dropped from 30.2% to 0.0%.
- Efficiency: Mean output tokens were 58 tokens shorter, indicating the model reaches and stops at the answer more efficiently, reducing truncation rates by 20 points.
Intended Use and Limitations
This model is primarily a research checkpoint for studying instruction-following adaptation. It is not intended as a general-purpose assistant. Evaluations have been limited to the MATH-500 proxy, and no assessments for IFEval, IFBench, safety, or multilingual capabilities have been conducted. The model may also exhibit hallucinations, such as incorrectly identifying its creator as OpenAI, inherited from its pretraining.