B1K models and experiments
Model checkpoints, training recipes, evaluation code, and reports from the training_models workspace. The included task-00 models target turning_on_radio in BEHAVIOR-1K.
Contents
| Folder | Contents |
|---|---|
| b1k_sft | Three B1K final inference checkpoints: human demonstrations, successful codegen, and a 50:50 mix; each at checkpoint 14999, with matching normalization and provenance. |
| pi05_task00_codegen_success_sft | Pi0.5 successful-codegen fine-tune, selected step-6000 EMA checkpoint, tokenizer, normalization, and reproduction source. |
| pi05_task00_sft | Training preparation, configuration, scripts, and receipts. |
| pi05_task00_sft_plan | Training plan and dataset audit. |
| pi05_tail_balanced_v2 | Tail-balanced training configuration, operational scripts, and tokenizer. |
| evaluation | Evaluation tools, results, and analyses. |
| reports | Reports and figures. |
| rollouts | Rollout-related scripts and supporting files. |
Download
pip install -U huggingface_hub
hf download davidwdw/b1k_model --local-dir ./b1k_model
To download only one model package:
hf download davidwdw/b1k_model --include 'b1k_sft/**' --local-dir ./b1k_model
# Or use --include 'pi05_task00_codegen_success_sft/**'.
These are custom JAX/Orbax OCDBT checkpoints. Keep each complete checkpoint tree intact and use its matching normalization, tokenizer, and inference implementation. See each package's documentation for setup, provenance, checksums, evaluation results, and limitations. Training or validation loss alone does not establish rollout success.
The original package documentation and receipts preserve historical local paths and source locations; adapt those paths to your installation. Virtual environments, runtime caches, Git internals, and temporary process files are omitted from this upload.