Modified Model: cognitivecomputations/Qwen3-30B-A3B-AWQ
Qwen models often experience language contamination during Arabic generation, particularly with less frequently trained languages. To address this, I conducted an experiment using Smoothie-Qwen to implement language suppression specifically during Arabic text generation.
Configuration
- Base model: cognitivecomputations/Qwen3-30B-A3B-AWQ
- Minimum scale factor: 0.5
- Smoothness: 10.0
- Sample size: 1000
- Window size: 4
- N-gram weights: [0.5, 0.3, 0.2]
Unicode Ranges
- Range 1: 0x4e00 - 0x9fff
- Range 2: 0x3400 - 0x4dbf
- Range 3: 0x20000 - 0x2a6df
- Range 4: 0xf900 - 0xfaff
- Range 5: 0x3040 - 0x309f
- Range 6: 0x30a0 - 0x30ff
- Range 7: 0x3100 - 0x312f
- Range 8: 0x3130 - 0x318f
- Range 9: 0xac00 - 0xd7af
- Range 10: 0x900 - 0x97f
- Range 11: 0x980 - 0x9ff
- Range 12: 0xa00 - 0xa7f
- Range 13: 0xa80 - 0xaff
- Range 14: 0xb00 - 0xb7f
- Range 15: 0xb80 - 0xbff
- Range 16: 0xc00 - 0xc7f
- Range 17: 0xc80 - 0xcff
- Range 18: 0xd00 - 0xd7f
- Range 19: 0xd80 - 0xdff
- Range 20: 0xe00 - 0xe7f
- Range 21: 0xe80 - 0xeff
- Range 22: 0xf00 - 0xfff
- Range 23: 0x1000 - 0x109f
- Range 24: 0x1780 - 0x17ff
- Range 25: 0x400 - 0x4ff
- Range 26: 0x500 - 0x52f
- Range 27: 0x370 - 0x3ff
- Range 28: 0x10a0 - 0x10ff
- Range 29: 0x590 - 0x5ff
- Range 30: 0x700 - 0x74f
- Range 31: 0x780 - 0x7bf
- Range 32: 0x800 - 0x83f
- Range 33: 0x1200 - 0x137f
- Range 34: 0x2d30 - 0x2d7f
- Range 35: 0x7c0 - 0x7ff
- Range 36: 0x13a0 - 0x13ff
- Range 37: 0x1400 - 0x167f
- Range 38: 0x1800 - 0x18af
- Range 39: 0xa000 - 0xa48f
- Range 40: 0x1950 - 0x197f
- Range 41: 0x16a0 - 0x16ff
- Range 42: 0x1680 - 0x169f
Statistics
- Target tokens: 41,835
- Broken tokens: 1,457
- Modified tokens: 43,179
- Downloads last month
- 8