TIR-Qwen3.5-4B-Thinking-v2

TIR-Qwen3.5-4B-Thinking-v2

Full-parameter SFT of Qwen/Qwen3.5-4B on the cleaned TIR mixture (HEART77/TIR-Training-Data, config thinking_clean, 26,515 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-Qwen3.5-4B-Thinking (trained on the legacy mixture). Compared with that run it drops the 2,200-row Cotrain attachment (260 images reused 10x each, fabricated "Simulated output" tool traces, 540 rows leaking the answer inside the reasoning) and corrects the synthetic color / refcoco chain-of-thought numerics by re-executing each shown code block on the real image. 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

Known limitations of this mixture for thinking-mode training, carried over from the legacy recipe: only 33% of rows contain a <think> block, the reasoning is short (median 141 characters for AgenticData, 525 for the synthetic set), and 4,000 AgenticData rows are bare single-answer responses without reasoning or code.

Training configuration

• Base model: Qwen/Qwen3.5-4B; full-parameter SFT, vision tower and aligner frozen
• Recipe: thinking (add_non_thinking_prefix=false, 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 = 414/414 optimizer steps, elapsed 3h 57m 44s, final training loss 0.14929664
• 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-Qwen3.5-4B-Thinking-v2 --port 8000 --max-model-len 32768 --reasoning-parser qwen3

The model uses the TIR agent format: assistant turns may open with <think>...</think>, emit Python in <code> blocks, receive results as a user turn beginning with Code execution result:, and finish with <answer>...</answer>.