Project notes, implementation writeups, and occasional research thoughts.
June 1, 2026
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.
May 25, 2026
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.
May 18, 2026
The tenth CodeSelf milestone adds the first agentic tool loop: public-test execution, revision, final submission, structured traces, and strategy metrics for self-debugging.
May 11, 2026
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.
May 4, 2026
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.
April 27, 2026
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.
April 20, 2026
The sixth CodeSelf milestone builds the baseline evaluation harness: pass@k, parser and reward diagnostics, informative prompt fraction, power analysis, and report generation.
April 13, 2026
The fifth CodeSelf milestone connects prompt rendering, model backends, code parsing, sandbox execution, reward scoring, and verbose rollout JSONL logs.
April 6, 2026
The fourth CodeSelf milestone turns sandbox execution results into an explainable correctness-first reward engine with structured diagnostics, normalization, and audit tooling.
March 30, 2026
The third CodeSelf milestone adds dataset loaders, hidden-test sidecars, leakage checks, deterministic splits, and fingerprints so benchmarks flow through one canonical task format.
March 23, 2026
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.
March 16, 2026
The first CodeSelf milestone focuses on research scaffolding before reinforcement learning: task schemas, config conventions, model headroom audits, and correctness-first rewards.
Posts will appear here after an HTML file is added under blog/posts/ and registered in data/blog.json.