Model feels pretty creative and has some nice moments. Couple brainfarts and bits of repetition occasionally, but nothing out of the normal. (The qwen team themselves recommend a presence penalty of 1.5. Yikes)

Non thinking and thinking are both supported. Thinking has reduced censorship as the original thinking refusals didn't seem to generalize well to the new format I gave it.

02 SillyTavern Settings

Recommended Roleplay Format

ActionsIn plaintext

Dialogue"In quotes"

Thoughts*In asterisks*

Recommended Samplers

Temp0.8

MinP0.05 - 0.075

Rep Pen1.00 - 1.1

Instruct

ChatML - Think

ChatML - NoThink

03 Quantizations

GGUF

iMatrix

04 Creation Process

Creation Process: SFT

SFT on approx 23 million tokens (12 million trainable). New is some Gemini Synth data which replaces some of my lower quality datasets.

About 10% of the dataset included reasoning for creative assistant tasks. This reasoning seems to have generalized quite well to other parts of the model and heavily reduces the token usage of thinking.

I think this model still needs a pass over with DPO to try and tackle the repetition and some of the weird oddities of the original instruct model, but that'll need to wait. I've been overspending training these models recently.

Trained using MS-Swift.

MS-Swift Config

SFT (4*H200)

PYTORCH_CUDA_ALLOC_CONF='expandable_segments:True' \
USE_HF=True \
WANDB_PROJECT=Qwen3.5-27B-SFT \
CUDA_VISIBLE_DEVICES=0,1,2,3 \
NPROC_PER_NODE=4 \
swift sft \
    --model Qwen/Qwen3.5-27B \
    --tuner_type lora \
    --dataset '/workspace/think_dataset.jsonl' \
              '/workspace/nothink_dataset.jsonl' \
    --torch_dtype bfloat16 \
    --bf16 true \
    --use_liger_kernel true \
    --lora_rank 128 \
    --lora_alpha 16 \
    --use_rslora true \
    --target_modules all-linear \
    --freeze_llm false \
    --freeze_vit true \
    --freeze_aligner true \
    --per_device_train_batch_size 1 \
    --gradient_accumulation_steps 16 \
    --num_train_epochs 2 \
    --learning_rate 2e-5 \
    --warmup_ratio 0.05 \
    --max_length 10752 \
    --split_dataset_ratio 0.01 \
    --add_non_thinking_prefix true \
    --load_from_cache_file true \
    --group_by_length true \
    --eval_steps 200 \
    --save_steps 200 \
    --save_total_limit 10 \
    --logging_steps 1 \
    --dataloader_num_workers 8 \
    --output_dir output/Qwen3.5-27B-SFT-Model \
    --report_to wandb

zerofata/Q3.5-BlueStar-27B

BlueStar v1

zerofata/
Q3.5-BlueStar-27B