Using the cold start dataset provided by Ares for SFT training, the MathVision evaluation score is only 29.

[gaozilve-max]

Thank you for your contribution; this is an outstanding piece of work. I encountered a slight issue during the cold start reproduction phase and would like to seek assistance.
My training configuration is as follows:
`### model
model_name_or_path: ./Qwen/Qwen2.5-VL-7B-Instruct
image_max_pixels: 2007040
video_max_pixels: 16384
trust_remote_code: true

method

stage: sft
do_train: true
finetuning_type: full
freeze_vision_tower: true
freeze_multi_modal_projector: true
freeze_language_model: false
deepspeed: ./LLaMA-Factory/examples/deepspeed/ds_z3_config.json

dataset

dataset_dir: ./huggingface.co/datasets/ares_sft
dataset: filter_data_final
template: qwen2_vl
cutoff_len: 32768
max_samples: null
overwrite_cache: true
preprocessing_num_workers: 16
dataloader_num_workers: 4

output

output_dir: ./checkpoints/llama_factory/ares_coldstart_big_image_filtered
logging_steps: 10
save_steps: 10
plot_loss: true
overwrite_output_dir: true
save_only_model: false
report_to: none

train

per_device_train_batch_size: 1
gradient_accumulation_steps: 8
learning_rate: 2.0e-5
num_train_epochs: 2
lr_scheduler_type: cosine
warmup_ratio: 0.05
bf16: true
gradient_checkpointing: true
ddp_timeout: 180000000
resume_from_checkpoint: null

special_tokens

add_tokens: ",,,"
skip_special_tokens: false
resize_vocab: true`

My training logs are as follows:

When I evaluated on the MathVision dataset, I used GPT-4o-mini for assessment and enabled the prefetch function. However, the accuracy rate was only 0.29, which shows a significant gap compared to the results in the original paper. I would like to ask if there might be any issues in my reproduction process.

I would be very grateful if you could help me resolve this problem.

[CSfufu]

Thanks a lot for trying ARES and for sharing such detailed configs and logs! 🙌

A score around 0.29 on MathVision typically indicates the cold-start SFT hasn’t trained sufficiently (or the effective batch / LR is too small). A few concrete suggestions that have helped us:

1. Increase effective batch size + LR

• Try an effective batch of ~256 tokens/sequences (e.g., per_device_train_batch_size: 4 with gradient_accumulation_steps: 64, or any equivalent that fits your GPUs).
• Set learning_rate: 2e-4 (cosine, warmup_ratio: 0.05 is fine).
• Train longer (e.g., 3–5 epochs) and watch for the loss to drop sharply in the first 1–2 epochs—if it doesn’t, it’s almost always undertraining.

2. Freezing policy

• For cold-start SFT, keeping freeze_language_model: false is good.
• You can keep the vision tower frozen but I’d recommend unfreezing the multimodal projector (freeze_multi_modal_projector: false) so it can adapt to the SFT data distribution.

3. Special tokens / template

• Since you’re using template: qwen2_vl, please make sure special tokens match that template. The current add_tokens: ",,," looks unusual—consider removing custom commas and relying on the template’s defaults (image/video placeholders, etc.). Keep resize_vocab: true.

4. Other knobs that help stability

• Maintain bf16: true, gradient_checkpointing: true, and your Deepspeed config.
• cutoff_len: 32768 is okay; just ensure your context packing (if any) doesn’t break multimodal alignment.
• Keep your image resolution cap reasonable (your image_max_pixels looks fine).

5. Evaluation protocol

• MathVision can be sensitive to prompting and grading. If possible, try our exact-match / rule-based evaluation once, or double-check the GPT-4o-mini grader rubric and the prefetch prompt format to ensure it mirrors our released eval script.

Minimal config tweaks (illustrative):

train:
  per_device_train_batch_size: 4
  gradient_accumulation_steps: 64
  learning_rate: 2.0e-4
  num_train_epochs: 4  # 3–5 usually converges
method:
  freeze_vision_tower: true
  freeze_multi_modal_projector: false
special_tokens:
  add_tokens: ""   # rely on qwen2_vl defaults
  resize_vocab: true

If you can share:

• your effective batch size (across all GPUs),
• a loss curve screenshot, and
• one short training log excerpt around the first few thousand steps,

I can help pinpoint whether it’s just undertraining or something token/template-related.

Thanks again for reporting — and great job pushing the reproduction forward! 🙏