hemantgaikwad06/aimo3-inference-engineering

How hemantgaikwad06/aimo3-inference-engineering works

This repository records an AIMO3 competition system and its engineering findings; it does not contain the runtime implementation in the supplied tree. The documented workflow runs from a mathematical problem through model inference and verification to an aggregated numerical answer. The graph reflects the README’s described design, not inspected notebook code; deployment and experiment branches are documented separately.

Main entry points: inference_pipeline.md Inference workflow · README.md Eight attempts · lessons_learned.md Stochastic variance · deployment_failures.md LoRA attempts

Components 16

Inference workflow

Harmony formatting README.md
Inference workflow docs/inference_pipeline.md
vLLM inference README.md
GPT-OSS-120B README.md
Eight attempts README.md
Deterministic anchor README.md

Verification and voting

Engineering findings

Stochastic variance docs/lessons_learned.md
Deployment constraints docs/deployment_failures.md
LoRA attempts docs/deployment_failures.md
Experiment findings experiments/notes.md

Other components

Problem source

Connections 14

Problem source → Harmony formatting: provides problem
Harmony formatting → Inference workflow: formats input
Inference workflow → vLLM inference: runs inference
vLLM inference → GPT-OSS-120B: queries model
Inference workflow → Eight attempts: generates attempts
Eight attempts → Deterministic anchor: includes anchor
Inference workflow → Python verification: verifies outputs
Python verification → Symbolic and numeric checks: uses checks
Python verification → Persistent Jupyter kernels: uses kernels
Inference workflow → Entropy-weighted voting: aggregates attempts
Entropy-weighted voting → Numerical answer: selects answer
Eight attempts → Stochastic variance: exposes variability
LoRA attempts → Deployment constraints: encounters constraints
Experiment findings → Stochastic variance: records findings

Drawn by GitDiagram on September 26, 2026 from the repository's file tree, README and source files. Read as Markdown ·

How hemantgaikwad06/aimo3-inference-engineering works

Connections 14

Problem source → Harmony formatting: provides problem
Harmony formatting → Inference workflow: formats input
Inference workflow → vLLM inference: runs inference
vLLM inference → GPT-OSS-120B: queries model
Inference workflow → Eight attempts: generates attempts
Eight attempts → Deterministic anchor: includes anchor
Inference workflow → Python verification: verifies outputs
Python verification → Symbolic and numeric checks: uses checks
Python verification → Persistent Jupyter kernels: uses kernels
Inference workflow → Entropy-weighted voting: aggregates attempts
Entropy-weighted voting → Numerical answer: selects answer
Eight attempts → Stochastic variance: exposes variability
LoRA attempts → Deployment constraints: encounters constraints
Experiment findings → Stochastic variance: records findings

Drawn by GitDiagram on September 26, 2026 from the repository's file tree, README and source files. Read as Markdown ·