branch	step	use
`main` = `step-1275`	last (1275/1275)	latest
`step-1200`	best eval_loss (0.317)	recommended
`step-1100`	1100
`step-300`	300	early (mode-collapse)
`step-200`	200
`step-100`	100	first working ckpt

Load a specific checkpoint:

python

from transformers import AutoModelForImageTextToText, AutoProcessor

model = AutoModelForImageTextToText.from_pretrained(
    "zhiyuanhucs/qwen3.5-4b-bc-sft-delta-force",
    revision="step-1200",                 # change to any branch above
    trust_remote_code=True,
    torch_dtype="bfloat16",
    device_map="auto",
)
processor = AutoProcessor.from_pretrained("zhiyuanhucs/qwen3.5-4b-bc-sft-delta-force", revision="step-1200", trust_remote_code=True)

Output format

Every assistant reply is a 200-ms action wrapped in custom tokens:

<|action_start|>X Y Z ; k1 ; k2 ; k3 ; k4 ; k5 ; k6<|action_end|>

X Y Z = cumulative mouse delta + wheel, quantized at 5-pixel step
6 key groups separated by ; (one per 33-ms sub-frame)
letter keys upper-case (W/A/D/Shift/Ctrl/LB/RB/MB/Q/F/...)

4 new special tokens added to tokenizer (vocab size 248320): <|action_start|>, <|action_end|>, <|thought_start|>, <|thought_end|>

Training recipe

Backend: Megatron-SWIFT (mcore-bridge 1.2.1, megatron-core 0.16.1)
Hardware: 4× H200 141GB
Parallelism: TP=4, PP=1, DP=1, Sequence Parallel=true
Batch: micro=1, global=8 (grad_acc=8)
Max len: 32768 (20 images × 1024 vision tokens + text)
LR: 1e-5 cosine → 1e-6, 5% warmup, 3 epochs
Weight decay: 0.01
Precision: bf16
Recompute: full / uniform / n=1
MTP: num_layers=1, scale=0.1 (Multi-Token Prediction, DeepSeek-V3 style)
Optimizer: AdamW (β1=0.9, β2=0.95)

Total training time: 6 h 27 min on 4× H200.

Evaluation

Format-correctness on 50 val samples (see scripts/42_check_format.py in the training repo):

Metric	step-1200
contains `<\|action_start\|>`	100%
contains `<\|action_end\|>`	100%
well-formed pair	100%
parses `X Y Z` as 3 ints	100%
uses `;` separator	100%
exact_match (body, after strip `<think></think>`)	14%
avg edit distance	6.2

Observation: keystrokes are learned correctly (semantic level — frequency of W/Shift/D/A/LB/RB/Q/F matches ground truth closely). Mouse delta magnitudes are still mode-collapsed toward ±10 — this is the expected limitation for a 4B model on dense-value regression. Use a 9B+ model for higher mouse precision.

Training code

Full training pipeline (dataset preprocessing, training scripts, checkpoint-format verification, scaling to 30B/70B) is at: https://github.com/game-agent-open/training