AnchorSR-Qwen3.5-9B-V4-RL-Step3000

AnchorSR-Qwen3.5-9B-V4-RL-Step3000

Intermediate V4 full-parameter GRPO checkpoint at step 3000 of a planned 7000-step run. Training was stopped after this checkpoint; this is not a completed one-epoch model.

Initialized from AnchorSR-Qwen3.5-9B-SFT-v4. Weights are exported losslessly to FP32 Safetensors from eight FSDP shards. Tokenizer, processor, and chat template are included; optimizer and training RNG states are excluded. Inference KV cache is enabled.

Training configuration

• 8 NVIDIA B300 GPUs; full-parameter GRPO.
• Global batch: 16 prompts, 4 responses per prompt (64 trajectories per step).
• Checkpoint: 3000 updates, 48,000 prompt presentations.
• Learning rate: 1e-6; KL loss coefficient: 0.02.
• Prompt/response limits: 8192/2048 tokens.
• V4 explicit video views with 16, 24, or 32 selected frames; preserve the corresponding view metadata when reproducing training inputs.
• Internal validation mean rule reward at step 3000: 0.4426009064 on 256 examples. This is not benchmark accuracy.

Loading

Use a Transformers version with Qwen3.5 support:

import torch
from transformers import AutoModelForImageTextToText, AutoProcessor

repo = "AnchorSR/AnchorSR-Qwen3.5-9B-V4-RL-Step3000"
processor = AutoProcessor.from_pretrained(repo)
model = AutoModelForImageTextToText.from_pretrained(
    repo, dtype=torch.bfloat16, device_map="auto"
)

Apply the included chat template and multimodal processor to inputs.