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
# captchaforge VLM training Python dependencies.
#
# Pin every version so `docker build` is reproducible across the
# trainer fleet. Bumps here are deliberate — re-test the full
# pipeline before committing a version delta.
# Core ML
torch==2.5.1
torchvision==0.20.1
transformers==4.46.3
accelerate==1.1.1
peft==0.13.2
trl==0.12.1
bitsandbytes==0.44.1
datasets==3.1.0
# Vision-language model deps
pillow==11.0.0
qwen-vl-utils==0.0.8
# Distillation extras
sentencepiece==0.2.0
protobuf==5.28.3
# HF Hub interop (model push/pull)
huggingface-hub==0.26.2
safetensors==0.4.5
# Auto-labelling teachers (OpenAI / Anthropic clients)
openai==1.54.4
anthropic==0.39.0
# Manifest parsing — match the captchaforge VlmDataset JSONL shape
orjson==3.10.11
# Bench-comparison side
numpy==1.26.4
pandas==2.2.3