srikarkashyap/iol-ai-2026-baseline
The srikarkashyap/iol-ai-2026-baseline is a 7.6 billion parameter instruction-tuned language model based on Qwen2.5-7B-Instruct, specifically adapted for the IOL-AI 2026 challenge. It utilizes 4-bit quantization (bitsandbytes NF4) for efficient inference and incorporates per-item prompting with step-by-step reasoning. This model is optimized for competitive AI challenges requiring structured problem-solving and answer extraction.
Loading preview...
IOL-AI 2026 Baseline (v3)
This model, srikarkashyap/iol-ai-2026-baseline, is a submission for the IOL-AI 2026 challenge. It is built upon the Qwen2.5-7B-Instruct base model, quantized to 4-bit using bitsandbytes NF4 for efficient deployment and inference within a constrained evaluation environment.
Key Adaptations and Features
- Quantization: Uses 4-bit (NF4) quantization for reduced memory footprint, with a fallback to fp16 if bitsandbytes is unavailable.
- Per-Item Prompting: Generates responses for each numbered item within a problem, rather than a single generation per problem.
- Structured Reasoning: Incorporates brief step-by-step reasoning, with the final answer extracted from a
FINAL:line. - Task-Specific Formatting: Includes task-type-specific instructions for answer formatting.
- One-Shot Example: Prompts include a one-shot worked example to guide the model's response generation.
- Time Budget Integration: Features a global time budget mechanism that dynamically adjusts reasoning space as the 30-minute limit approaches.
- Self-Contained: Model weights are shipped within the repository to ensure functionality in evaluation sandboxes without internet access.
Use Cases
This model is specifically designed and optimized for:
- Participating in the IOL-AI 2026 challenge.
- Structured problem-solving tasks requiring step-by-step reasoning and precise answer extraction.
- Environments with strict resource constraints where 4-bit quantized models are beneficial.