Repository navigation
Tags
Tags
Release 2026-11-05 Batched changes: - Use Inkling 2 Small in human-facing documentation - [Release] Register Inkling-2-Small metadata, measured LoRA counts, and release guidance - [Release] Support H200k rendering for Inkling-2-Small - Migrate to Tinker SDK 0.38 unified API futures - Skip unused sampler exports during async drain - Add rollout UI design and verification skill - Stream async training and late evaluation metrics as independent events - Add training-experiment defaults to AGENTS.md
Release 2026-10-27 Batched changes: - Give every recipe a default Tinker session name - Drop the transformers upper bound and exclude known-bad releases - Add num_epochs to RL training and warn when max_steps exceeds it - Evaluate the final weights in the sync RL loops - Use a dedicated content tokenizer in Qwen3 renderers - Train the final batches when streaming-minibatch async training shuts down
Add Inkling-Small to hyperparam lookup tables (thinking-machines-lab#930 ) Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
PreviousNext