StudentSim/configs/training/tutor_sft.yaml at main · microsoft/StudentSim · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
# Supervised finetuning for the chess tutor.
#
# The tutor reads a rendered board alongside the text, so this is the one
# training run in the repository that is multimodal. The vision tower stays
# frozen: the corpus teaches what to say about a position, not how to see one.

base_model: Qwen/Qwen3-VL-8B-Instruct
stage: 1

data_path: tutor_rl/sft_corpus.jsonl
output_dir: tutor_rl/sft

# 1 per device x 32 accumulation x 8 GPUs. Board images make the per-device
# batch the binding constraint, so the accumulation carries the batch size.
per_device_batch: 1
gradient_accumulation: 32
world_size: 8
effective_batch: 256

max_seq_len: 2048
precision: bf16
gradient_checkpointing: false

freeze_vision_tower: true
max_pixels: 200704          # 448 x 448

epochs: 1.0
save_freq: 200
save_total_limit: 3

trainer_seed: 42
data_sampler_seed: 42

lora:
  rank: 32
  alpha: 64
  dropout: 0.05
  target_modules:
    - q_proj
    - k_proj
    - v_proj
    - o_proj
    - gate_proj
    - up_proj
    - down_proj

optimizer:
  name: adamw_torch_fused
  learning_rate: 1.0e-4
  beta1: 0.9
  beta2: 0.95
  weight_decay: 0.1
  grad_clip: 1.0
  schedule: cosine
  warmup_steps: 20