GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
nassimjp
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
nassimjp
Model Tree
Qwen/Qwen3.5-9B
Input Modalities
Output Modalities
Supported Functionality
This model was structured and optimized 2x faster using Unsloth to complete the native tokenization space for the Pashto language.
The original Qwen3.5 tokenizer natively supports 42 Pashto alphabet characters as single unique tokens. However, 7 critical Pashto-specific characters were missing from its native vocabulary, forcing the tokenizer to fallback into split, multi-token byte sequences (Byte-level BPE fragmentation).
This precision patch initializes and injects only those 7 missing characters into the vocabulary via the Unsloth wrapper. This approach ensures 100% native Pashto alphabet coverage without disrupting preexisting multi-lingual merge rules or vocabulary weights.
text
ټ ډ ږ ڼ څ ځ ښ| Tokenizer Profile | Native Pashto Characters Supported | Missing Native Alphabet entries | Evaluation Status |
|---|---|---|---|
| Original Qwen3.5 | 42 | 7 | ⚠️ Subword Fragmentation |
| Qwen3.5-Pashto-Base (Ours) | 49 | 0 | ✅ 100% Alphabet Integrity |
Each of the 7 newly added characters now strictly processes into exactly 1 unified token instead of leaking into multi-byte arrays:
ټ → 1 token (ID 248077)ډ → 1 token (ID 248078)ږ → 1 token (ID 248079)ڼ → 1 token (ID 248080)څ 1 token (ID )We evaluated token consumption using a standard native Pashto sample phrase:
text
"زه نن ښوونځي ته ځم ځکه چې هلته کتابونه لولم."28 tokens (Bloated due to byte fallbacks like Ú and ģ)24 tokens−4 tokens (~14.3% sequence compression efficiency)This is a Base Model intended for Continued Pretraining (CPT) or fine-tuning workflows on massive Pashto corpora. To pull this repository and target the new embeddings for down-stream training, initialize the following Unsloth framework:
python
import torchfrom unsloth import FastLanguageModel model_name = "nassimjp/qwen3.5-9b-4bit-pashto-base"max_seq_length = 4096 model, tokenizer = FastLanguageModel.from_pretrained( model_name = model_name, max_seq_length = max_seq_length, dtype = torch.bfloat16, # Optimized for Ampere/Ada Lovelace architectures load_in_4bit = True,) # Configure LoRA adapters while explicitly saving the new token weightsmodel = FastLanguageModel.get_peft_model( model, r = 32, lora_alpha = 32, target_modules = ["q_proj", "k_proj", "v_proj", "o_proj", "gate_proj", "up_proj", "down_proj"], lora_dropout = 0, bias = "none", modules_to_save = ["embed_tokens", "lm_head"], # Crucial for freezing/saving the 7 injected characters)This repository hosts a foundational Base Model, not a conversational assistant or instruction-tuned checkpoint. It requires Supervised Fine-Tuning (SFT) or Reinforcement Learning from Human Feedback (RLHF) before handling chat applications.
248081ځ → 1 token (ID 248082)ښ → 1 token (ID 248083)