apoorva2311/Limopro-model

qwen_limopro_output

This model is a fine-tuned version of Qwen/Qwen2.5-3B-Instruct on the acereason_limopro_15k dataset.

Model description

Qwen-LIMOPro-Output builds on Qwen2.5-3B-Instruct, a general-purpose instruction-following model.
Through supervised fine-tuning (SFT) on the AceReason-LIMOPro subset, the model was optimized to improve mathematical and logical reasoning.

The subset emphasizes:

• Challenging problems (high reasoning complexity)
• Unfamiliar examples (low similarity to base model’s prior knowledge)
• Diverse coverage across topics (via K-means clustering of embeddings)

This approach helps strengthen reasoning depth while minimizing overfitting and catastrophic forgetting often seen in small-scale SFT.

Intended uses & limitations

Use Cases

• Step-by-step mathematical reasoning
• Problem-solving and scientific question answering
• Educational or research applications for reasoning-focused LLMs

Limitations

• Not suitable for general open-domain QA or creative text generation (model over-specialized toward reasoning tasks)
• Some degradation observed on general benchmarks (e.g., MMLU)
• May over-explain or generate verbose reasoning chains for simple prompts

Training and evaluation data

• Training Dataset: apoorva2311/TML-limopro

  • 15K examples sampled from AceReason-1.1-SFT (100K total)
  • Selected via difficulty and model familiarity scoring
  • Ensures high reasoning quality and topic diversity

• Evaluation Benchmarks:

  • AIME24 / AIME25
  • MATH-500
  • GPQA-Diamond
  • LiveCodeBench (LCB)
  • MMLU-Redux-2

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 5e-06
• train_batch_size: 1
• eval_batch_size: 8
• seed: 42
• distributed_type: multi-GPU
• num_devices: 4
• gradient_accumulation_steps: 16
• total_train_batch_size: 64
• total_eval_batch_size: 32
• optimizer: Use adamw_torch with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: cosine
• lr_scheduler_warmup_ratio: 0.05
• num_epochs: 3

Average Results (vs. Baseline Qwen2.5-3B-Instruct)

Benchmark	Baseline	Random SFT	LIMOPro SFT
MATH-500	0.6340	0.5060	0.5600
AIME25	0.0000	0.0333	0.1000
MMLU (Acc)	0.6393	0.2356	0.4404

The LIMOPro variant achieved +22% higher overall accuracy compared to random fine-tuning, with improved stability and reasoning alignment.

Training results

Training Configuration

• Environment: Virginia Tech ARC (Tinkercliffs, 4×L40S GPUs)
• Framework: LLaMA-Factory
• Fine-tuning Type: Full-parameter SFT
• Precision: BF16
• Deepspeed Config: ds_z3_offload_config.json

Hyperparameters

Parameter	Value
Learning Rate	5e-6
Batch Size	1
Gradient Accumulation	16
Total Batch Size	64
Epochs	3
Scheduler	Cosine
Warmup Ratio	0.05
Optimizer	AdamW (β₁=0.9, β₂=0.999, ε=1e-8)
Seed	42

---

• The model exhibits stronger structured reasoning and clearer intermediate step articulation.
• Slightly reduced performance on broad factual tasks (trade-off for reasoning specialization).
• Fine-tuning on the LIMOPro-selected subset led to smoother loss curves and better convergence stability compared to random sampling.

Framework versions

• Transformers 4.56.2
• Pytorch 2.7.1+cu126
• Datasets 4.0.0
• Tokenizers 0.22.1