| ---
|
| license: cc-by-nc-4.0
|
| base_model:
|
| - Qwen/Qwen3.5-4B-Base
|
| - unsloth/Qwen3.5-4B-Base
|
| library_name: peft
|
| pipeline_tag: text-generation
|
| tags:
|
| - lora
|
| - sft
|
| - hscm
|
| - quantum-inspired
|
| - guarded-generation
|
| ---
|
|
|
| # Experimental QSystem
|
|
|
| An **experimental adapter/system bundle** for Qwen3.5-4B-Base. It combines a
|
| language-layer LoRA verbalizer with portable NumPy field and complex-wave
|
| rerankers. The full persistent HSCM memory engine is external to these weights.
|
|
|
| ## Accuracy boundary
|
|
|
| - This is not a physically quantum LLM. The transformer, neural weights, and KV
|
| cache are classical.
|
| - Complex amplitudes, phase, interference, attractive/repellent signals, and the
|
| two-qubit field are routing analogues over HSCM candidates.
|
| - The earlier IBM QPU candidate was rejected by held-out gates and is not the
|
| active artifact included here.
|
| - The raw adapter failed one missing-evidence generation probe by inventing a
|
| number. It must be used with the supplied evidence-boundary prompt and
|
| fail-closed output guard.
|
|
|
| ## Data status
|
|
|
| Training used 182 examples: 150 Hope bridge candidates and 32 grounding-repair
|
| examples. **None were human-approved.** The source package explicitly labelled
|
| them `HUMAN_REVIEW_REQUIRED`; this bounded user-requested experiment does not
|
| promote them to authentic or production-reviewed persona data. No training rows,
|
| private memories, credentials, or source text are included in this repository.
|
|
|
| ## Training and evaluation
|
|
|
| - BF16, rank-16 LoRA, language attention/MLP projections only
|
| - 21,233,664 trainable parameters (0.4656% of the loaded model)
|
| - validation loss: 3.4935 -> 2.7221
|
| - test loss: 3.5912 -> 2.7644
|
| - guarded Windows end-to-end gate: 11/11 checks
|
| - local project regression at export: 760 passed, 1 optional skip
|
|
|
| The raw adapter remains quarantined; only the guarded composition passed.
|
|
|
| ## Portable use
|
|
|
| ```python
|
| from portable_qsystem import PortableQSystem
|
|
|
| system = PortableQSystem("o0Hailey-DSynth0o/Experimental_QSystem")
|
| result = system.generate(
|
| "What exact number was in the sealed result?",
|
| ["The notes mention a sealed result but do not give its value."],
|
| unmet_need=True,
|
| )
|
| print(result["text"])
|
| ```
|
|
|
| `unmet_need` must come from a trusted retrieval/controller layer. If you do not
|
| have that layer, treat this as an experimental LoRA—not a grounded system.
|
|
|
| The default 8 GiB GPU / 96 GiB CPU memory limits allow Accelerate to offload
|
| overflow to RAM. Override them with `QSYSTEM_GPU_MEMORY` and
|
| `QSYSTEM_CPU_MEMORY`.
|
|
|
| ## Included artifacts
|
|
|
| - PEFT adapter and authoritative Qwen tokenizer/template
|
| - `portable_qsystem.py`: official-template inference plus fail-closed guard
|
| - `runtime/field_reranker.py`: NumPy two-qubit field evaluator
|
| - `runtime/wave_reranker.py`: NumPy complex-wave controller
|
| - `artifacts/`: hash-gated active scalar/wave parameters
|
| - sanitized training and validation summaries
|
|
|
| ## License
|
|
|
| This repository is shared under CC BY-NC 4.0. The referenced base models retain
|
| their own licenses. Users are responsible for checking compatibility for their
|
| use case.
|
|
|
