[DOC]: Clarify More about Datasets for Chat Training Stage 3

[JThh]

📚 The doc issue

As raised by some users, the current guide to training Stages 3 regarding the required datasets is yet clear. Some user has composed a guide which still confuses the datasets.

Some clarification is needed in this section, regarding how to download and set prompt_path and pretrain_dataset.

Below is an example user query I received on Slack.

torchrun --standalone --nproc_per_node=4 train_prompts.py \
         --pretrain "bigscience/bloom-560m" \
         --model 'bloom' \
         --strategy colossalai_zero2 \
         --prompt_path /data/chenhao/train/ColossalAI/prompt_dataset/data.json \   -------Where is the data.json data
         --pretrain_dataset /data/chenhao/train/ColossalAI/pretrain_dataset/data.json \    -----------Where is the data.json data
         --rm_pretrain /data/chenhao/train/ColossalAI/Coati-7B \
         --rm_path /data/chenhao/train/ColossalAI/rmstatic.pt \
         --train_batch_size 4 \
         --experience_batch_size 4 \
         --max_epochs 1 \
         --num_episodes 1