JITAI Qwen3 experimental training project — no trained weights
This repository shares code, validation schemas, 300 authored synthetic candidates,
and a Google Colab QLoRA smoke-test notebook targeting Qwen/Qwen3-1.7B.
No trained LoRA adapter, merged model, tokenizer export, or training results are
included. Local inspection found no adapter weights/configuration or saved
training report, including inside the project's three ZIP archives. A separate
Colab run outside this project cannot be verified from these local files.
This repository cannot be loaded as a trained model with from_pretrained.
Data and intended use
All 300 candidates remain synthetic and unreviewed. Automated validation is not human approval. Stage B has not been accepted. Splits contain 198 training, 52 validation, and 50 test candidates. The smoke test selects six training and six separate validation examples, one per decision class. Test examples are validated for dataset integrity but never fitted or scored by the smoke test. No claims about clinical safety, model quality, or generalization are supported. Use for research and teammate inspection; this is not a production or medical system. Host applications must independently enforce consent, safety, and tool execution. Safety/security examples include synthetic distress and adversarial requests.
Prepared experiment
The current notebook is configured for an explicitly acknowledged unreviewed-data experiment: eight optimizer steps, seed 20261015, 4-bit NF4 QLoRA, rank 8, alpha 16, q_proj/v_proj targets, and assistant-only loss. Read the actual notebook for the sequence cap. It rejects overlength examples rather than truncating. If a sequence check fails, no weight download or training proceeds; inspect the failure in Colab rather than treating a prepared notebook as a completed run. The base model commit is resolved and recorded at execution time; no executed base revision or GPU environment can be claimed here. GPU determinism is not guaranteed. The source contains no executed training evidence or measured training metrics.
Reproduce and collaborate
See REPRODUCING.md for CPU validation, Colab steps, and artifact handling. Review source changes with teammates through Hub discussions and pull requests. This public repository is owned by the personal account AnuCaro. Teammates can propose changes through Hub pull requests; organization roles do not apply to this personal repository. No credentials, personal Drive contents, human-review working files, environments, model caches, or submission snapshots are part of this release.
The input ZIP is preserved byte-for-byte to match its preconfigured checksum. Its embedded notebooks/README are historical; use the top-level colab/stage_c_qwen3_experimental.ipynb and these reproduction instructions. Historical dataset-card warnings remain intact: experimental acknowledgement does not approve the dataset for an approved pilot or production use.
License and provenance
The existing project has no explicit project/data license. This release does not invent one; the team must decide a license before granting broader reuse rights. The base model has its own terms at https://huggingface.co/Qwen/Qwen3-1.7B. No base-model weights are redistributed. See PUBLICATION_STATUS.json for the local artifact inventory and SHA256SUMS for hashes of the release files.