smithtune v0.1.4

Adds configurable validation intervals and more precise checkpoint selection for Fireworks Training API and Baseten Loops. (#123)

• Set --validation-every-steps N to validate at optimizer-step intervals as well as epoch ends. Epoch-end validation remains the default.
• Select the lowest-loss checkpoint across all validation points, with its epoch and global step recorded for replay.
• See steps per epoch and expected validation passes in training plans.
• Complete the requested epochs by default; enable optional epoch-based early stopping with --early-stopping-patience N.
• Preserve completed validation results incrementally for manual recovery.
• Group provider code and tests into provider-specific folders.

Upgrade

uv tool install --force --python 3.12 \
  --overrides https://raw.githubusercontent.com/langchain-ai/smithtune/v0.1.4/overrides.txt \
  'smithtune[deepagents] @ git+https://github.com/langchain-ai/smithtune.git@v0.1.4'

For Baseten deployment, retain the extra by installing smithtune[deepagents,baseten-deploy] from the same tag.

Full changes: v0.1.3...v0.1.4

smithtune v0.1.3

Adds dedicated Fireworks training and detached training controllers on existing remote machines.

• Run Fireworks LoRA training on dedicated capacity with automatic shape selection or a pinned --training-shape. Dedicated replay manages temporary base and tuned serving capacity. (#119)
• Use train --detach on an existing Linux VM for workflows requiring an active controller, then inspect progress with status and logs. Runs continue after your local machine disconnects; controller failures require manual recovery using available artifacts and checkpoints. (#120)
• Consult the new Training optimizations reference during planning for context sizing, batching, and throughput/cost comparisons. (#122)

Upgrade

uv tool install --force --python 3.12 \
  --overrides https://raw.githubusercontent.com/langchain-ai/smithtune/v0.1.3/overrides.txt \
  'smithtune[deepagents] @ git+https://github.com/langchain-ai/smithtune.git@v0.1.3'

If you use Baseten deployment, retain that extra by installing smithtune[deepagents,baseten-deploy] from the same tag.

Full changes: v0.1.2...v0.1.3

smithtune v0.1.2

Add Fireworks training support for GLM-5.3-Flash with configurable reasoning effort.

• Train GLM-5.3-Flash through Fireworks with a maximum context of 200,064 tokens. (#117)
• Choose prepare --reasoning-effort low|high|max (default: max). SmithTune saves the choice for training and base-versus-tuned replay. All three levels keep thinking enabled; --reasoning-policy separately controls recorded source reasoning. (#117)
• Keep generated GLM reasoning separate from visible replay answers. (#117)
• Clarify evaluator feedback selection in the packaged workflow skill. (#114)
• Update the urllib3 dependency lock to 2.8.0 and pin GitHub Actions to commit SHAs. (#115, #113)
• Keep model-specific configuration details in the skill, CLI help, and preparation reference. (#118)

Upgrade

uv tool install --force --python 3.12 \
  --overrides https://raw.githubusercontent.com/langchain-ai/smithtune/v0.1.2/overrides.txt \
  'smithtune[deepagents] @ git+https://github.com/langchain-ai/smithtune.git@v0.1.2'

If you use Baseten deployment, retain that extra by installing smithtune[deepagents,baseten-deploy] from the same tag.

Full changes: v0.1.1...v0.1.2

smithtune v0.1.1

Patch release with improvements to dataset recovery, replay evaluation, and deployment.

• Skip trajectories with foreign trace evidence instead of aborting the entire dataset pull; report the exclusion and retain none of that trajectory's messages or tool bindings. (#97)
• Resume triage with different concurrency or retry attempts while retaining completed votes. (#98)
• Accept dictionary tool arguments during replay scoring and show safe, actionable request-failure summaries with resume guidance. (#104)
• Surface Baseten provisioning failures. (#101)
• Automatically select Fireworks deployment shapes for the requested account, wait for a ready replica, and bound creation and readiness polling with one timeout. Preserve deployment details and cleanup guidance after startup timeouts. (#99)

Upgrade

uv tool install --force --python 3.12 \
  --overrides https://raw.githubusercontent.com/langchain-ai/smithtune/v0.1.1/overrides.txt \
  'smithtune[deepagents] @ git+https://github.com/langchain-ai/smithtune.git@v0.1.1'

If you also use Baseten deployment, retain that extra by installing smithtune[deepagents,baseten-deploy] from the same tag.

Existing completed replay scores are not automatically recomputed. Use a fresh --output-dir to rescore evaluations affected by dictionary tool arguments.

Full changes: v0.1.0...v0.1.1