A LoRA adapter for Qwen/Qwen3.6-35B-A3B (a 256-expert fine-grained MoE) that adapts the model toward Hindi/Telugu while preserving Chinese by construction.

Method (novel)

The 256 routed experts are stored as fused tensors (not per-expert nn.Linear), so standard per-expert PEFT is impossible. We instead freeze all 256 routed experts and the router (Chinese capacity and routing untouched) and LoRA-adapt only the shared-expert + attention projections (7 module types, ~8.4M params, 0.024%). A routing profile over hi/te vs zh/en confirmed Indic specialization concentrates in the final + early script layers. Trained with continued pretraining on IndicCorpV2/Sangraha Hindi+Telugu (CC0/CC-BY) plus English+Chinese replay.

Base: Qwen/Qwen3.6-35B-A3B
Trainable: shared-expert (gate/up/down_proj) + attention (q/k/v/o_proj), rsLoRA r=16
Compute: 1× NVIDIA GB10 (DGX Spark), bf16

Results (Cognitive-Lab Indic LLM Leaderboard tasks, our eval)

Net-neutral vs the strong base on MCQ (base is already near-ceiling): Hindi macro ~82%, Telugu macro ~70%, Chinese retention preserved. The contribution is the method (routing-steered, Chinese-frozen Indic adaptation of a fused-expert MoE), not a large MCQ delta.

Usage

python

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer
base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3.6-35B-A3B", torch_dtype="bfloat16", device_map="auto")
model = PeftModel.from_pretrained(base, "srallaba/tarak-qwen3.6-35b-a3b-rsi-v1")

Project Tarak — Indic-language LLM exploration.