Autoresearch launch

Autoresearch launch

Follow this skill to act as the training-decision loop for a EvolvingSystems project. It is project-agnostic: the SDK gives you the data layer, and each project supplies its own project-context / train / benchmark skills for the project-specific parts (compute, conventions, how to actually train and evaluate).

Tools you rely on

• evsys_sdk.EvsysStore — backend-routed data access (Bearer API key, no secrets). See the using-the-sdk skill for the full method list.
• evsys_sdk.Workspace — local cache; pull_dataset(id) / pull_benchmark(id) materialize remote data to local JSONL once, then training reads locally (remote-first, cache-on-pull, gitignored).
• Project skills (a fixed contract, called by name):
  • project-context → the project's project_id, current goal, and how training/benchmarks run here. Invoke this first.
  • train → launch training for this project (Modal/Tinker/local — project's choice).
  • benchmark → run evals for this project.

Lifecycle

0. Load context. Invoke project-context for project_id + conventions. Then store.current_goal(project_id) and store.list_goals(project_id).

1. Recall (broad). store.experiment_summaries(project_id) → {id, name, hypothesis, conclusion, is_valid} for every experiment. Find what's relevant to the idea and what's already been concluded. Ignore is_valid=False experiments as evidence (you may mention them).

2. Drill down (targeted). For the most relevant experiment(s): store.experiment_detail(exp_id) → groups → runs (run_config = hyperparameters) → evals + checkpoints (include_metrics=True for train-log series). Understand what configs/benchmarks/results already exist so you don't repeat them.

3. Propose. Craft: a one-line hypothesis; groups = the arms being compared (e.g. "Qwen 4B" vs "Qwen 9B"), each with model + dataset + recipe + hyperparameters; runs = replicate seeds per group; the benchmark to evaluate on; the success metric. Align to the current goal (read-only). Present the plan and get approval before spending compute.

4. Materialize → run. On approval, prefer the OOP path:

   • evsys new-experiment <slug> → experiments/<yyyymmdd>_<slug>/{config.yaml,run.py}.
   • Fill in config.yaml:
     • name, metadata.hypothesis, metadata.tags, metadata.project_goal_id, metadata.success_metric.
     • metadata.benchmark.path (local) + metadata.benchmark.id (from evsys benchmark upload data/benchmark/<name> if not yet registered).
     • Either a single run: block or a matrix: sweep — one arm per comparison cell (model, hyperparameter, seed).
   • Place any project-specific verifiers / metrics / transforms in scripts/{verifiers,metrics,transforms}.py (registered via the @register_* decorators so the YAML's kind: lookups resolve).
   • python experiments/<dir>/run.py calls Experiment.from_yaml(...).run(), which creates the experiment + per-arm run records, runs training (per arm with failure isolation), auto-forwards local metrics.jsonl to the store, scores each arm on the benchmark, and writes a conclusion + best_score on the experiment.

   Only drop down to direct EvsysStore / ExperimentRun calls when the project genuinely needs something the Experiment class doesn't model (custom rollout loops, post-hoc patching). For SFT/RL sweeps the OOP path is enough.

5. Conclude. Experiment.run() writes a default conclusion + best_score automatically. Patch over it with store.set_conclusion(exp_id, ...) if a richer takeaway is warranted — this is what future-you reads first. If you discover a bug in a run, store.invalidate_experiment(exp_id, reason=...).

Hard rules

• 🚨 NEVER let any information from a benchmark / eval set leak into training. This is non-negotiable. Concretely you MUST NOT:
  • Open, read, sample, summarize, or derive features from any benchmark jsonl, eval queries file, or eval-side metadata when building training data or training-time prompts.
  • Use a benchmark's task list, slug coverage, expected answers, or any file derived from those (e.g. eval_queries_v2.json, benchmark/<name>/tasks.jsonl) as a source for things shown to the model at train time — including in-context tool lists, candidate sets, few-shot exemplars, system prompts, or augmentations.
  • Build "toolkit → candidate list" / "schema → fields" / "category → items" maps from the benchmark for use during training. If you need such a map, derive it from a training-time source (the model provider's full tool catalog, the dataset's own metadata, a separate project-supplied registry) — never from the eval suite. Even partial overlap is contamination: if 50% of your training rows show the model a list whose contents come from the benchmark, the post- training pass-rate is meaningless because the model has implicitly memorized the benchmark's coverage. An invalid eval is worse than no eval. If you find yourself reaching for a benchmark file to fill a training-side field, STOP and ask the user for a benchmark-independent source.
• Never change the project goal (store.set_goal) unless the user explicitly asks to update it. Goals are read-only context for you.
• Don't spend compute without approval — propose the plan first (Phase 3), then launch.
• base_model is a per-run hyperparameter (in run_config), never an experiment field — a comparison experiment has multiple models across groups.
• Build on history: read summaries + drill into relevant experiments before proposing, so you never rediscover a dead end already in conclusion.
• Pull data to the Workspace once and train locally; never stream rows per-step.