https://ardxa.com/blog.html Project notes, research notes, and occasional writing by Harsit Upadhya. en-us https://ardxa.com/blog/posts/milestone-12-publishable-methodology-package.html https://harsitupadhya.com/blog/posts/milestone-12-publishable-methodology-package.html Mon, 01 Jun 2026 12:00:00 GMT The twelfth CodeSelf milestone turns the project into a reproducible research package with methodology docs, checklists, model and dataset card templates, limitations, and manifest tooling. https://ardxa.com/blog/posts/milestone-11-ppo-comparison-scaffolding.html https://harsitupadhya.com/blog/posts/milestone-11-ppo-comparison-scaffolding.html Mon, 25 May 2026 12:00:00 GMT The eleventh CodeSelf milestone adds a PPO comparison scaffold with a dependency-free PPO smoke trainer, diagnostics, rollout-budget checks, and PPO-versus-GRPO reports. https://ardxa.com/blog/posts/milestone-10-agentic-strategy-extension.html https://harsitupadhya.com/blog/posts/milestone-10-agentic-strategy-extension.html Mon, 18 May 2026 12:00:00 GMT The tenth CodeSelf milestone adds the first agentic tool loop: public-test execution, revision, final submission, structured traces, and strategy metrics for self-debugging. https://ardxa.com/blog/posts/milestone-09-final-held-out-evaluation-and-statistics.html https://harsitupadhya.com/blog/posts/milestone-09-final-held-out-evaluation-and-statistics.html Mon, 11 May 2026 12:00:00 GMT The ninth CodeSelf milestone adds final held-out evaluation tooling with paired task outcomes, exact McNemar-style testing, bootstrap confidence intervals, task flips, and inspectable comparison reports. https://ardxa.com/blog/posts/milestone-08-main-grpo-experiment-orchestration.html https://harsitupadhya.com/blog/posts/milestone-08-main-grpo-experiment-orchestration.html Mon, 04 May 2026 12:00:00 GMT The eighth CodeSelf milestone turns GRPO smoke-test scaffolding into an experiment matrix for seeds, group sizes, KL betas, reward versions, curriculum modes, summaries, and checkpoint selection. https://ardxa.com/blog/posts/milestone-07-grpo-smoke-test-scaffolding.html https://harsitupadhya.com/blog/posts/milestone-07-grpo-smoke-test-scaffolding.html Mon, 27 Apr 2026 12:00:00 GMT The seventh CodeSelf milestone scaffolds a dependency-free GRPO smoke trainer to validate grouped rollouts, rewards, advantages, informative prompts, metrics, and checkpoint manifests before real fine-tuning. https://ardxa.com/blog/posts/milestone-06-baseline-evaluation.html https://harsitupadhya.com/blog/posts/milestone-06-baseline-evaluation.html Mon, 20 Apr 2026 12:00:00 GMT The sixth CodeSelf milestone builds the baseline evaluation harness: pass@k, parser and reward diagnostics, informative prompt fraction, power analysis, and report generation. https://ardxa.com/blog/posts/milestone-05-base-model-inference-and-parsing.html https://harsitupadhya.com/blog/posts/milestone-05-base-model-inference-and-parsing.html Mon, 13 Apr 2026 12:00:00 GMT The fifth CodeSelf milestone connects prompt rendering, model backends, code parsing, sandbox execution, reward scoring, and verbose rollout JSONL logs. https://ardxa.com/blog/posts/milestone-04-reward-engine.html https://harsitupadhya.com/blog/posts/milestone-04-reward-engine.html Mon, 06 Apr 2026 12:00:00 GMT The fourth CodeSelf milestone turns sandbox execution results into an explainable correctness-first reward engine with structured diagnostics, normalization, and audit tooling. https://ardxa.com/blog/posts/milestone-03-dataset-loaders-and-hidden-tests.html https://harsitupadhya.com/blog/posts/milestone-03-dataset-loaders-and-hidden-tests.html Mon, 30 Mar 2026 12:00:00 GMT The third CodeSelf milestone adds dataset loaders, hidden-test sidecars, leakage checks, deterministic splits, and fingerprints so benchmarks flow through one canonical task format. https://ardxa.com/blog/posts/milestone-02-sandboxed-code-execution.html https://harsitupadhya.com/blog/posts/milestone-02-sandboxed-code-execution.html Mon, 23 Mar 2026 12:00:00 GMT The second CodeSelf milestone builds the execution layer: syntax checks, static safety scans, subprocess execution, hidden-test handling, and phase-level results for future rewards. https://ardxa.com/blog/posts/milestone-01-project-scaffold-and-research-baseline.html https://harsitupadhya.com/blog/posts/milestone-01-project-scaffold-and-research-baseline.html Mon, 16 Mar 2026 12:00:00 GMT The first CodeSelf milestone focuses on research scaffolding before reinforcement learning: task schemas, config conventions, model headroom audits, and correctness-first rewards.