docs: add detailed dataset composition and methodology
Browse files
README.md
CHANGED
|
@@ -103,6 +103,40 @@ $$\mathcal{L}_{\text{RLCD}} = \mathcal{L}_{\text{CE}} + 0.4 \mathcal{L}_{\text{B
|
|
| 103 |
|
| 104 |
---
|
| 105 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 106 |
## 🇰🇷 한국어 안내 (Korean Overview)
|
| 107 |
|
| 108 |
### Qwev-9B-RLCD란?
|
|
|
|
| 103 |
|
| 104 |
---
|
| 105 |
|
| 106 |
+
## 📂 Training Data Composition & Construction Methodology
|
| 107 |
+
|
| 108 |
+
The model was trained on a meticulously curated **5-in-1 Decision Alignment Mixture** (4,800 primary + 1,700 STEM expansion records) designed to prevent shortcut learning and enforce mathematically sound probability distributions:
|
| 109 |
+
|
| 110 |
+
| Dataset Component | Source | Samples | Key Function & Calibration Objective |
|
| 111 |
+
| :--- | :--- | :---: | :--- |
|
| 112 |
+
| **Enterprise Typed Decisions** | `LocalLLaMA/typed-decisions` | 1,800 | Multi-criteria enterprise routing, workflow state parsing, and API dispatching. |
|
| 113 |
+
| **Human Preference Judges** | `allenai/reward-bench` | 1,000 | Direct pairwise preference alignment ($P(\text{chosen}) > P(\text{rejected})$). |
|
| 114 |
+
| **Evidence-Deficient Uncertainty** | `kev-suites / boolq` | 1,000 | Ground-truth stripped contexts enforcing uniform $1/K$ entropy regularization. |
|
| 115 |
+
| **Long-Context Needle Attention** | Synthetic Needle Retrieval | 500 | 1k~3k token noise contexts training pointer survival across long sequences. |
|
| 116 |
+
| **Ambiguous Soft-Target NLI** | `alisawuffles/WANLI` | 500 | Continuous non-binary soft target probabilities for subtle semantic boundaries. |
|
| 117 |
+
| **10-Choice STEM Reasoning** | `TIGER-Lab/MMLU-Pro` | 1,700 | 10-way choice options (`opt_A` to `opt_J`) preventing softmax collapse on high-cardinality decisions. |
|
| 118 |
+
|
| 119 |
+
### Data Formatting Format:
|
| 120 |
+
Every sample is represented as a structured JSON decision record:
|
| 121 |
+
```json
|
| 122 |
+
{
|
| 123 |
+
"state": "Context / problem description or dialogue transcript",
|
| 124 |
+
"questions": {
|
| 125 |
+
"q0": {
|
| 126 |
+
"type": "choice",
|
| 127 |
+
"instructions": "Select the factually accurate or superior decision.",
|
| 128 |
+
"criteria": {
|
| 129 |
+
"opt_A": "First candidate option description",
|
| 130 |
+
"opt_B": "Second candidate option description"
|
| 131 |
+
},
|
| 132 |
+
"label": "opt_A"
|
| 133 |
+
}
|
| 134 |
+
}
|
| 135 |
+
}
|
| 136 |
+
```
|
| 137 |
+
|
| 138 |
+
---
|
| 139 |
+
|
| 140 |
## 🇰🇷 한국어 안내 (Korean Overview)
|
| 141 |
|
| 142 |
### Qwev-9B-RLCD란?
|