Upload ARO Coder 4-bit (distill_student)
Browse files
README.md
CHANGED
|
@@ -26,8 +26,8 @@ ARO is a domain-specific language where every statement follows the pattern:
|
|
| 26 |
| **Base model** | [mlx-community/Qwen3-Coder-30B-A3B-Instruct-4bit](https://huggingface.co/mlx-community/Qwen3-Coder-30B-A3B-Instruct-4bit) |
|
| 27 |
| **Quantization** | 4-bit (MLX) |
|
| 28 |
| **Language** | ARO |
|
| 29 |
-
| **Training samples** |
|
| 30 |
-
| **Syntax pass rate** |
|
| 31 |
| **Source label** | distill_student |
|
| 32 |
|
| 33 |
## Links
|
|
@@ -108,7 +108,7 @@ Key features:
|
|
| 108 |
|
| 109 |
This model was trained with the ARO training pipeline:
|
| 110 |
|
| 111 |
-
1. **Corpus collection** —
|
| 112 |
2. **Supervised fine-tuning** — LoRA on all code generation, debugging, Q&A, and explanation tasks
|
| 113 |
3. **DPO preference training** — using `aro check` validation to build chosen/rejected pairs
|
| 114 |
4. **Iterative self-improvement** — multiple rounds of generate-validate-retrain
|
|
|
|
| 26 |
| **Base model** | [mlx-community/Qwen3-Coder-30B-A3B-Instruct-4bit](https://huggingface.co/mlx-community/Qwen3-Coder-30B-A3B-Instruct-4bit) |
|
| 27 |
| **Quantization** | 4-bit (MLX) |
|
| 28 |
| **Language** | ARO |
|
| 29 |
+
| **Training samples** | 3630 |
|
| 30 |
+
| **Syntax pass rate** | 58% |
|
| 31 |
| **Source label** | distill_student |
|
| 32 |
|
| 33 |
## Links
|
|
|
|
| 108 |
|
| 109 |
This model was trained with the ARO training pipeline:
|
| 110 |
|
| 111 |
+
1. **Corpus collection** — 3630 samples from Examples, Book, Wiki, Proposals, and real-world ARO applications
|
| 112 |
2. **Supervised fine-tuning** — LoRA on all code generation, debugging, Q&A, and explanation tasks
|
| 113 |
3. **DPO preference training** — using `aro check` validation to build chosen/rejected pairs
|
| 114 |
4. **Iterative self-improvement** — multiple rounds of generate-validate-retrain
|