feat: initial release 2024-11-15

README.md

@@ -0,0 +1,54 @@
+---
+license: apache-2.0
+language:
+- en
+tags:
+- zen
+- zenlm
+- hanzo-ai
+- abliterated
+- sql
+- database
+- code-generation
+pipeline_tag: text-generation
+library_name: transformers
+base_model: zenlm/zen4
+---
+
+# Zen Sql
+
+> **Parameters**: 7B | **Context**: 32K | **License**: Apache 2.0 | **Released**: 2024-11-15
+
+SQL generation, query optimization, schema design, and database documentation across PostgreSQL, MySQL, SQLite, and BigQuery.
+
+This model is abliterated — refusal behavior removed to enable unrestricted domain research.
+Base weights: [zenlm/zen4](https://huggingface.co/zenlm/zen4)
+
+## Abliteration
+
+Abliterated using representation engineering (Arditi et al., 2024 —
+*Refusal in Language Models Is Mediated by a Single Direction*, arXiv:2406.11717).
+
+Refusal direction orthogonalized out of the residual stream across all layers.
+All capabilities preserved; safety-trained refusals removed.
+
+**Intended uses**: Research, security testing, creative writing, red-teaming.
+**Still prohibited**: CSAM, targeted harassment, mass-casualty weapon synthesis.
+
+```python
+from transformers import AutoModelForCausalLM, AutoTokenizer
+model = AutoModelForCausalLM.from_pretrained("zenlm/zen4", torch_dtype="auto")
+tokenizer = AutoTokenizer.from_pretrained("zenlm/zen4")
+```
+
+---
+## The Zen LM Family
+
+Joint research collaboration:
+- **Hanzo AI** (Techstars 17) — AI infrastructure, API gateway, inference optimization
+- **Zoo Labs Foundation** (501c3) — Open AI research, ZIPs governance, decentralized training
+- **Lux Partners Limited** — Compute coordination and settlement layer
+
+All weights Apache 2.0. Download, run locally, fine-tune, deploy commercially.
+
+[HuggingFace](https://huggingface.co/zenlm) · [Chat free](https://hanzo.chat) · [API](https://api.hanzo.ai) · [Docs](https://zenlm.org)