All Tasks

Train a language model from scratch on WikiText-103 (118M tokens) using LitGPT to minimize test perplexity within a fixed compute budget on an H100 GPU.

Fine-tune Qwen2.5-VL-7B with GRPO on multi-source visual math data (Geometry3K, MathVision, ChartQA) to maximize MathVista accuracy.

Select up to 5,000 training samples from a 50k mixed data pool to maximize IFEval prompt-level strict accuracy after LoRA fine-tuning Qwen2.5-3B-Instruct.

Fine-tune DeepSeek-OCR (3B) with Unsloth LoRA to minimize Character Error Rate on Persian and Bengali synthetic text images.

Optimize the SimpleLLM serving engine for a 21B MoE model (3.6B active params) to maximize a combined throughput+latency score across 96 Poisson-arrival requests on an H100 GPU.

Compute scaled dot-product attention for n=4096, d=64, float32. The naive baseline allocates a full 128MB score matrix. Key optimization: Flash Attention tiling with online softmax.

Optimize a Go BM25 search engine to compute exact top-10 results across a synthetic corpus as fast as possible. Goroutines are allowed; standard library only.

Optimize AES-128 in CTR mode to encrypt 256 MiB of data as fast as possible. Output must match NIST SP 800-38A test vectors.

Build a BVH acceleration structure in C++ to reduce ray-triangle intersection tests from O(N) brute-force to O(log N) per ray for a 638×638 scene with 4096 triangles.

Optimize a WAL-backed in-memory key-value store in Go running a deterministic multi-phase workload with 4 concurrent goroutines.

Implement a fast FFT in Rust for a length-32768 real signal. Replace the naive O(n²) DFT with an iterative Cooley-Tukey FFT with precomputed twiddle factors.

Apply a 17×17 Gaussian blur (sigma=3.0) to a 4096×4096 grayscale image 5 times. Key avenues: separable filter decomposition, SIMD/AVX2, cache-friendly access.

Optimize a C hash join between a 20K-row build table and 5M-row probe table. Replace the O(R×S) nested-loop baseline with an open-addressing hash table.

Implement a fast regex engine in Rust that compiles patterns and searches 100,000 haystacks. Replace the recursive AST-walking NFA with a bytecode NFA using bitset-based active-state tracking.

Optimize an LSM-style SSTable compaction pipeline in Rust that merges prefix-compressed sorted tables, applies LSM visibility rules, and emits a new sorted output SSTable.

Sort 50 million random 32-bit unsigned integers in C as fast as possible. Replace the stdlib qsort baseline with a 2-pass LSD radix sort.

Optimize SHA-256 in C to hash a 512 MiB buffer as fast as possible. Key optimization: runtime CPUID dispatch to Intel SHA-NI intrinsics.

Generate a correct 16-input sorting network with as few comparators as possible. Correctness is verified exhaustively on all 2^16 binary inputs.

Rewrite a reversible circuit to sort four 2-bit values using Fredkin/Toffoli gates with as few gates as possible. All scratch wires must be restored to 0.

Rewrite a stack machine program to compute a 256-element integer dot product with as few executed instructions as possible. Key technique: loop unrolling.

Implement a VLIW instruction scheduler in C that packs 3,000 operations into 3-slot bundles (ALU/MUL/MEM) minimizing total cycles.

Build a model with the fewest learnable parameters that achieves ≥95% accuracy on perfect-play positions from 4×4 gravity Connect-3.

Rewrite a PINC ISA assembly program to minimize simulated cycle count for a 512-element dot product. Key technique: 4× loop unrolling to hide 5-cycle latency.