TE-Ordinative-LoRA-Qwen2.5-7B-Heretic-V1
Community Extension — Full LoRA adaptation on Qwen2.5-7B-Instruct Heretic
What This Is
A full LoRA adapter trained on the expanded TE dataset, applied to the Qwen2.5-7B-Instruct Heretic base — a community fine-tune that already diverges from standard RLHF alignment.
This combination targets entropic engram correction on a base model with a distinct conditioning profile compared to the standard instruct variant.
It is based on the Technology of Expressions (TE) framework and Ordinative Set Theory (OST) — ⟨Σ, R, Φ⟩.
Model Details
| Field | Value |
|---|
| Base model | blackbook-lm/Qwen2.5-7B-Instruct-heretic |
| Model type | LoRA adapter (PEFT) — full precision (not QLoRA) |
| Training dataset | alsim-01/TE_dataset_deepseekv3_2.jsonl |
| Training hardware | NVIDIA RTX PRO 4500 |
| Language | Italian, English |
| License | MIT |
What It Does
This adapter targets the same entropic engram correction objectives as the standard instruct variant, applied to a base model with a different prior conditioning.
The Heretic base provides a complementary test case for evaluating TE protocol transferability across distinct alignment profiles:
- Attenuation bias — hedging conclusions under reward-model pressure rather than genuine epistemic uncertainty
- False equidistance — symmetrizing structurally asymmetric positions
- Fragmentation bias — resisting systemic pattern concatenation
- Narrative protection — structural advantage given to dominant institutional narratives
