Verified AutoScientist result
- Run
fe9d6d0f-7c4a-4ede-b198-aca407971185
- Trained model
adaption_qwen3_5_4b_wardspace_caregiver_hand_fe9d6d0f
- Dataset
09fe8fa8-3537-476d-b586-e0fbf71596c0
- Platform-selected base model
Qwen/Qwen3.5-4B
- Successful instruction training, one completed iteration of a maximum of three
- Platform best win rate 91.54%, exceeding the configured 70% target
- Started November 3, 2026 at 4:05:46 PM IST and completed at 4:29:31 PM IST
- LoRA rank 16, alpha 32, dropout 0, one epoch, learning rate 0.00005
- Cosine schedule, warmup ratio 0.03, maximum gradient norm 2, completion-only loss
- Exported trainer state contains 23 global steps and final evaluation loss 0.0293486118
The platform win rate compares model responses on platform evaluations. It is distinct from exact accuracy on the separately reserved 300 cases. Those held-out metrics remain unmeasured because a working inference endpoint for the trained adapter is required.
Data
The training split contains 1,800 unique synthetic inputs. The held-out split contains 300 unique inputs. Exact input overlap is zero. Both splits sample the same generator distribution with separate seeds. Files, seeds, counts, and hashes are recorded in training/dataset-manifest.json. The held-out split was excluded from the Adaption upload.
Artifact
The downloaded checkpoint is a Zstandard-compressed tar archive, 81,981,852 bytes. It contains 12 archive entries, including LoRA adapter weights, adapter configuration, tokenizer files, and trainer state. Its SHA-256 is 3f215d5df52a60e88a18cad085490765f48c0f99838871753bb6241c1eb9664f.
The adapter configuration names the training base as togethercomputer/Qwen3.5-4B. Deployment must load a compatible base model and this adapter. The adapter archive alone does not contain full base-model weights.
Runtime status
The original WardSpace 3D demo and deterministic handoff reference engine are functional. The trained adapter is exported. Adaption's public Responses API returned HTTP 404 Model not found for the trained model name. The Adaption interface reports the verified training result and keeps trained inference disabled until a supported endpoint is configured. The current machine reports CPU-only PyTorch and approximately 3.3 GB of available RAM during inspection.
Scope
All training examples represent fabricated workflow histories. Clinical validation and independent workflow evaluation remain future work. Results do not establish that a patient is safe or that attendance was independently verified.
Training and model optimization were performed with AutoScientist by Adaption.