Replace incorrect README with accurate description

README.md

@@ -1,45 +1,48 @@
 ---
 license: apache-2.0
-pipeline_tag: text-generation
-tags:
-  - code-generation
-  - scientific-computing
-  - agentic
-  - npcpy
-  - no-comments
 ---
 
 # enpisi-coder
 
-An agentic scientific coding model trained to write clean, self-documenting code with no comments and no unnecessary try/except blocks. Trained on scientific computing tasks across physics, astronomy, AI/ML, and data analysis.
+LoRA adapter for [npcsh](https://github.com/NPC-Worldwide/npcsh) agentic command execution.
 
-## Design principles
+Trained on npcsh benchmark traces to improve tool calling, shell command generation, and file operation reasoning in the Qwen3 family of models.
 
-- **No comments** — code is self-documenting through clear naming. Comments are noise.
-- **No defensive try/except** — unless explicitly requested. Errors should propagate loudly so they can be debugged, not silently swallowed.
-- **Agentic workflow** — the model searches, computes, verifies, then writes. It doesn't guess.
-- **Scientific precision** — trained on real scientific computing tasks (CHSH parameters, BPT classifiers, Bayesian inference, etc.)
+## Files
 
-## Training
+| File | Purpose |
+|------|---------|
+| `adapters.safetensors` | LoRA weights |
+| `adapter_config.json` | Metadata (base model, rank, alpha) |
 
-- **SFT**: Style-constrained coding traces generated via [npcpy](https://github.com/NPC-Worldwide/npcpy)
-- **RL (DPO)**: Reward function penalizes comments (-0.15/line) and unnecessary try/except (-0.30/block), rewards verification steps and compilable code
-- **Data sources**: 77M word corpus spanning physics, astronomy, AI/ML, philosophy, and literature — used to generate diverse Q&A and coding training pairs
+## Base Model
+
+- **Qwen3-4B** (`mlx-community/Qwen3-4B-4bit`)
 
 ## Usage
 
+### With npcsh (MLX)
+
+```bash
+export NPCSH_CHAT_MODEL="mlx-community/Qwen3-4B-4bit"
+export NPCSH_CHAT_PROVIDER="omlx"
+# Place adapter folder at ~/.npcsh/adapters/
+```
+
+### With transformers (Python)
+
 ```python
 from npcpy.ft.sft import load_sft_model, predict_sft
 
 model, tokenizer = load_sft_model("npc-worldwide/enpisi-coder")
 response = predict_sft(
     model, tokenizer,
-    "Implement a BPT diagram classifier for emission line galaxies",
+    "List all Python files in this directory",
     max_new_tokens=500
 )
 ```
 
-## Part of the NPC Worldwide ecosystem
+## Ecosystem
 
-- [npcpy](https://github.com/NPC-Worldwide/npcpy)
-- [npcsh](https://github.com/NPC-Worldwide/npcsh)
+- [npcpy](https://github.com/NPC-Worldwide/npcpy) — training toolkit
+- [npcsh](https://github.com/NPC-Worldwide/npcsh) — NPC shell