Clean Sky-3B-Q model card
Browse files
README.md
CHANGED
|
@@ -62,25 +62,6 @@ The main fine-tuning source was:
|
|
| 62 |
|
| 63 |
- `WithinUsAI/claude_mythos_distilled_25k`
|
| 64 |
|
| 65 |
-
The run used a cleaned subset and small project-specific anchors:
|
| 66 |
-
|
| 67 |
-
| Source / domain | Rows used |
|
| 68 |
-
|---|---:|
|
| 69 |
-
| Math fallback anchors | 910 |
|
| 70 |
-
| Advanced coding | 595 |
|
| 71 |
-
| Mathematical reasoning | 252 |
|
| 72 |
-
| Scientific analysis | 108 |
|
| 73 |
-
| Agentic planning | 45 |
|
| 74 |
-
| General expert QA | 27 |
|
| 75 |
-
| Sky identity | 16 |
|
| 76 |
-
| General knowledge | 10 |
|
| 77 |
-
| Coding anchors | 9 |
|
| 78 |
-
| Constitutional safety | 8 |
|
| 79 |
-
| Research notes | 6 |
|
| 80 |
-
| General no-identity-leak anchors | 5 |
|
| 81 |
-
| Writing anchors | 5 |
|
| 82 |
-
| Planning anchors | 4 |
|
| 83 |
-
|
| 84 |
The dataset cleaning removed or softened repeated source-branding strings so the model answers as **Sky**, not as Claude/Mythos. The identity set was intentionally kept small to avoid identity fixation.
|
| 85 |
|
| 86 |
## Quick Start: Google Colab
|
|
@@ -226,26 +207,3 @@ It is best used as an experimental open assistant checkpoint, not as a productio
|
|
| 226 |
## Safety
|
| 227 |
|
| 228 |
The fine-tune includes a small constitutional safety slice, but this is not a complete safety training process. Do not rely on it for high-risk domains without additional evaluation, guardrails, and monitoring.
|
| 229 |
-
|
| 230 |
-
## Citation / Credit
|
| 231 |
-
|
| 232 |
-
Base model:
|
| 233 |
-
|
| 234 |
-
```text
|
| 235 |
-
sapientinc/HRM-Text-1B
|
| 236 |
-
```
|
| 237 |
-
|
| 238 |
-
Main dataset:
|
| 239 |
-
|
| 240 |
-
```text
|
| 241 |
-
WithinUsAI/claude_mythos_distilled_25k
|
| 242 |
-
```
|
| 243 |
-
|
| 244 |
-
Project identity:
|
| 245 |
-
|
| 246 |
-
```text
|
| 247 |
-
Sky, made by the 0labs and CognixAI team.
|
| 248 |
-
Made in India.
|
| 249 |
-
CEO: Atharvsinh Jadav.
|
| 250 |
-
```
|
| 251 |
-
|
|
|
|
| 62 |
|
| 63 |
- `WithinUsAI/claude_mythos_distilled_25k`
|
| 64 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 65 |
The dataset cleaning removed or softened repeated source-branding strings so the model answers as **Sky**, not as Claude/Mythos. The identity set was intentionally kept small to avoid identity fixation.
|
| 66 |
|
| 67 |
## Quick Start: Google Colab
|
|
|
|
| 207 |
## Safety
|
| 208 |
|
| 209 |
The fine-tune includes a small constitutional safety slice, but this is not a complete safety training process. Do not rely on it for high-risk domains without additional evaluation, guardrails, and monitoring.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|