Qwen3.5-VL-2B — Proactive Agent SFT (AgentNet)
Supervised fine-tune of Qwen/Qwen3.5-2B on the agentnet_sft split of
KeithXi/Qwen3.5_Proactive_SFT, for proactive GUI-agent behaviour (the model
learns when to act and when to ask the user a clarifying question).
Three checkpoints from the same run at different points of the epoch:
| Directory | Step | Fraction of the epoch |
|---|---|---|
iter_7200/ |
7,200 | 54% |
iter_9623/ |
9,623 | 72% |
iter_13450/ |
13,450 | 100% (one full epoch) |
Training
| Base model | Qwen/Qwen3.5-2B |
| Data | KeithXi/Qwen3.5_Proactive_SFT, agentnet_sft, revision eb883526 |
| Samples | 26,900 of 26,901 rows, one epoch |
| Steps | 13,450 |
| Global batch size | 2 |
| Sequence length | 25,472 |
| Learning rate | 1e-6, constant |
| Parallelism | TP 2 x PP 2 on 4x H200 |
| Precision | bf16 |
| Trainable | language model only; the vision tower and the projector stay frozen |
| Framework | Megatron-Bridge, exported from torch_dist with AutoBridge.export_ckpt |
The upstream corpus is xlangai/AgentNet; conversations keep five turns of
history, at most twenty images, and relative coordinates. 839 of the samples
are ask_user targets.
Usage
from transformers import AutoModelForCausalLM, AutoProcessor
repo = "WFRaain/qwen35vl-2b-agentnet-sft"
model = AutoModelForCausalLM.from_pretrained(repo, subfolder="iter_13450", dtype="bfloat16")
processor = AutoProcessor.from_pretrained(repo, subfolder="iter_13450")
Notes
Weights only — the optimizer and RNG state live in the original Megatron
torch_dist checkpoints, which are not published here. No evaluation has been
run on these checkpoints yet.
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support