AAAAI – Qwen 3.6 Training

Qwen 3.6 Pipeline

Canonical weights: Qwen/Qwen3.6-35B-A3B
Use this lane for the staged flow you asked for: continue pretrain -> instruction SFT -> preference tuning (DPO/GRPO) -> eval + distill. The in-app runner handles continued pretrain, SFT, and GRPO. DPO and full distillation recipes are generated below for external stacks.

Repo: QwenLM/Qwen3.6 HF: Qwen/Qwen3.6-35B-A3B Default lane: AGI-POC + Unsloth

Pipeline Stage

Preference Method

Recipe Backend

Pipeline Notes

Generated Recipe

Auto-Create Dataset

Convert source material into a starter SFT dataset without leaving the page.

Upload PDF, CSV, DOCX or TXT

Model Configuration

Modality

Model

Custom Model ID

Training Type

Max Seq Length

HF Token

Hyperparameters

Epochs

Learning Rate

Batch Size

Grad Accumulation

Max Steps (0=auto)

Warmup Steps

Optimizer

LR Scheduler

Dataset Packing

Dataset

HuggingFace

Local File

Upload

Created

Data Source

HuggingFace Dataset

Subset

Train Split

Format

Eval Split

LoRA Settings

LoRA Rank (r)

LoRA Alpha

Target Modules

RS-LoRA

Reinforcement Learning (GRPO)

Enable GRPO

Training Loss

Learning Rate

Gradient Norm

Eval Loss

Training Loss

Learning Rate

Gradient Norm

Eval Loss

Loss (MA-20)

GPU Utilization

Loading hardware info...

Trained Models

ID	Base	Type	Date
Loading...

Export Model

Model

Format

GGUF Quantization

HuggingFace Hub

Hub Repo

HF Token

Export Reference

Format	Use Case
GGUF Q4_K_M	llama.cpp, Ollama, LM Studio
GGUF Q8_0	High quality CPU inference
Merged 16-bit	Full precision safetensors
LoRA only	Adapter weights (smallest)
UD-Q5_K_XL	Unsloth Dynamic (best ratio)

Test Model

Model

System Prompt

Prompt

Max Tokens

Temperature

Output

Run inference to see output...

Training Log

...