MLJAR Studio Changelog

Version 1.0.3

March 23, 2026

Stronger runtime stability, AutoLab data setup, and smarter run summaries

MLJAR Studio 1.0.3 focuses on reliability and flow quality across conversational notebooks and AutoLab experiments. This release adds stronger timeout/cleanup guards, improves experiment data and results-directory UX, introduces LLM-based run summaries, and fixes multiple state, session, and startup edge cases.

Back to changelog

Highlights

Added Ollama Cloud support for remote model connectivity
Added watchdog and timeout guards to prevent hanging "Thinking" and "Checking packages" states
Improved AutoLab Step 2 data setup with desktop-friendly multi-file selection and persisted `data_files`
Upgraded run summary generation to LLM-based synthesis and standardized report delivery via `run_summary`
Fixed notebook rename conflicts and prevented cross-notebook variable leakage in conversational workflows

Added

  • Added Ollama Cloud support for connecting to remote Ollama-hosted model endpoints.
  • Added Step 2 two-column AutoLab data setup UX with data specification on the left and file selection on the right.
  • Added desktop-compatible AutoLab file picker behavior with persistent selected-file list.
  • Added data_files persistence in experiment metadata/config and wired it into draft/run payloads.
  • Added selected-file context into AGENTS draft prompts for clearer data-path context.
  • Added proactive runtime access refresh (startup, focus, periodic) and stronger websocket re-initialization after close.
  • Added/standardized run summary delivery through report payload field run_summary.
  • Added LLM-based final run summary synthesis using notebook outcomes, recent events, error memory, and leaderboard context.
  • Added writing of run summary to artifacts/run_summary.md with final_summary.md kept for compatibility.
  • Added best-so-far outline to runtime metric chart (cumulative max/min by optimization direction).
  • Added lightweight chart legend and clearer marker styling for raw vs best-so-far interpretation.
  • Added Data Profile Step 2 target resolution (LLM pass for target/task/confidence/reason/alignment).
  • Added deterministic target-profile resolver with concrete stats (missingness, uniqueness, samples, balance/regression stats).
  • Added focused tests for new Data Profile flow (test_data_profile_step2.py), including example-form wiring and inventory truncation rules.

Improved

  • Improved conversational notebook stability with watchdog and timeout guards so long-running states do not hang indefinitely.
  • Improved package precheck/install flow with stronger cleanup, probe/install timeouts, and safer failure handling.
  • Improved notebook title rename reliability with conflict fallback (name.ipynb -> name_2.ipynb, etc.).
  • Improved session-expired UX by unfreezing chat input on send/auth errors and tightening reconnect flow.
  • Improved Results directory UX with inline validation messaging and Use suggested name quick action.
  • Improved example-preset first-run flow with automatic suggested results directory suffixing.
  • Improved run summary panel rendering/styling to avoid duplicate headings and improve containment.
  • Improved runtime chart tooltip content to show both raw and best-so-far values.
  • Simplified AGENTS draft profile context to a single source: data_profile_summary_md (summary.md only).
  • Updated Data Profile summary rendering with explicit LLM TARGET PROFILE section and metric alignment.
  • Updated Data Profile API to accept scoring form fields (score_name, score_compute) in both Step 1 and Step 2 prompts.
  • Unified Data Profile UX wording to one progress message: Generating data profile....

Fixed

  • Fixed cross-notebook variable leakage by attaching variables only after local execution context is established.
  • Fixed experiment startup crash with Ollama/runtime settings by avoiding deep-copy of lock-bearing runtime objects.
  • Fixed run-state reliability on startup failures by writing run_failed events and persisting failure reason in run state.json.
  • Fixed report inference to mark startup/early-failure runs as failed instead of falling back to Ready to start.
  • Changed results directory validation endpoint to read-only behavior; directories are now created only on Save & Run.