TIR-Qwen3.5-9B-NonThinking-v2

TIR-Qwen3.5-9B-NonThinking-v2

Full-parameter SFT of Qwen/Qwen3.5-9B on the cleaned TIR mixture (HEART77/TIR-Training-Data, config v1_nonthink_clean, 18,120 rows). One contiguous epoch at a strict 32,768-token maximum length, trained 2026-09-02/03 on 4xA100 80GB with ms-swift 4.5.2 and Megatron Core 0.16.1.

This is the successor of HEART77/TIR-Qwen3.5-9B-NonThinking (trained on the legacy mixture). Compared with that run it drops the 2,200-row Cotrain attachment (10x image reuse, direct answers without reasoning), 17 rows whose final turn lacked <answer> tags, and one row whose 196-megapixel intermediate image made ms-swift skip the sample. All other hyper-parameters are identical, so the two checkpoints form a data-only A/B pair. See the dataset card for the full audit.

Training data

No assistant message contains a <think> block; the empty non-thinking prefix is inserted by the template and excluded from the loss.

Training configuration

• Base model: Qwen/Qwen3.5-9B; full-parameter SFT, vision tower and aligner frozen
• Recipe: non-thinking (add_non_thinking_prefix=true, loss_scale=default+ignore_empty_think, truncation_strategy=delete)
• Learning rate 1e-5, min 1e-6, cosine decay, warmup fraction 0.05, weight decay 0.1, clip grad 1.0
• Maximum length 32,768; global batch 64; micro batch 1; TP=2, PP=1, sequence parallel; bf16; flash attention
• One epoch = 283/283 optimizer steps, elapsed 8h 10m 55s, final training loss 0.17167911
• Full arguments: training/args.json; per-step log: training/logging.jsonl

Evaluation

Not yet evaluated at publication time. Results on TIR-Bench and the huawei_benchmark suite (FVQA, SimpleVQA, MMBrowseComp, GAIA) will be added alongside the legacy checkpoint for comparison.

Inference

vllm serve HEART77/TIR-Qwen3.5-9B-NonThinking-v2 --port 8000 --max-model-len 32768 --reasoning-parser qwen3

Disable thinking at request time with chat_template_kwargs={"enable_thinking": false}. The model uses the TIR agent format: Python in <code> blocks, tool results as a user turn beginning with Code execution result:, final answer in <answer>...</answer>.