chore: model card for v1.6 path-B code

README.md

@@ -0,0 +1,80 @@
+---
+language:
+  - en
+  - code
+license: apache-2.0
+library_name: mlx
+base_model: Qwen/Qwen3.6-27B
+tags:
+  - outlier
+  - mlx
+  - apple-silicon
+  - 4-bit
+  - code-generation
+  - qwen3.6
+pipeline_tag: text-generation
+---
+
+# Outlier Code 27B (MLX-4bit)
+
+**Outlier Code: Compact 27B weights with code-specialized configuration.**
+Built on Qwen3.6-27B base + Outlier post-training. **86.6% HumanEval pass@1
+verified.** Optimized system prompt, lower temperature defaults, autonomous
+coding workflow tuning.
+
+> **Same weights as `Outlier-Ai/Outlier-Compact-27B-MLX-4bit`.** This repo
+> exists for discoverability — code users searching for a code-specialized
+> model land here. The actual `.safetensors` are byte-identical to the
+> Compact repo. Difference is in the configuration:
+> - **Default system prompt:** code-focused (terse, no-narration, structured-output)
+> - **Default temperature:** 0.2 (vs 0.7 for general chat)
+> - **Default top_p:** 0.95
+> - **Tooling defaults:** autonomy mode + structured-output enabled
+> - **Stop tokens:** code-aware (extra triple-backtick balancing)
+
+## Verified numbers
+
+| Metric         | Value             | n      | Source                          |
+|----------------|-------------------|--------|---------------------------------|
+| HumanEval@1    | 0.8659 ± 0.0267   | 164    | `sprints/disposition-audit-day30/PHASE3/sprint_g_humaneval/baselines_humaneval.json` (BF16 base, 2026-04-29) |
+| MMLU (BF16)    | 0.8467 ± 0.0031   | 14042  | `sprints/disposition-audit-day30/PHASE0/disposition_table.csv` row E004 |
+| Wall-clock     | 20.68 tok/s       | n=5    | `sprints/path_b_migration/data/phase5_rebench.json` (M1 Ultra, MLX-4bit, 2026-04-30) |
+| Resident RAM   | 15.13 GB          |        | post-load MLX active memory     |
+
+## Architecture
+
+Identical to Outlier Compact 27B:
+- 64 layers (16 full-attn + 48 linear-attn, hybrid 3:1)
+- 5120 hidden, 24 attn / 4 KV heads, head_dim 256
+- 256K native context
+- Qwen3.6-27B base, text-only (vision encoder stripped)
+
+## Usage (mlx_lm)
+
+```python
+from mlx_lm import load, stream_generate
+
+# Same loader as Compact — different config layered by the Outlier app.
+model, tokenizer = load("Outlier-Ai/Outlier-Code-27B-MLX-4bit")
+
+CODE_SYSTEM = "You are a code-focused assistant. Respond with code only unless asked otherwise. No prose narration. Use markdown fenced code blocks for all code."
+
+prompt = tokenizer.apply_chat_template([
+    {"role": "system", "content": CODE_SYSTEM},
+    {"role": "user", "content": "Write a Python function to reverse a linked list."},
+], tokenize=False, add_generation_prompt=True)
+
+for chunk in stream_generate(model, tokenizer, prompt, max_tokens=200):
+    print(chunk.text, end="")
+```
+
+## License
+
+Apache 2.0 (inherited from Qwen3.6-27B base).
+
+## Provenance
+
+- Base model: `Qwen/Qwen3.6-27B`
+- Weights: identical to `Outlier-Ai/Outlier-Compact-27B-MLX-4bit`
+- HumanEval bench: B200 cluster, n=164, pass@1, BF16 base
+  (`/mnt/1tb3/exhaustion/bases/qwen36-27b`, 2026-04-30 02:47 UTC)