Supervised fine-tuning (SFT) checkpoints for the CRAFT project, spanning four model scales (0.5B / 1.5B / 3B / 7B) built on the Qwen2.5-Instruct family. These serve as the SFT initialization for downstream GRPO / reward-shaping experiments.

Repository layout

Each checkpoint is stored as a directly-loadable subfolder named <scale>_<version>:

Scale	Base model	Subfolders
0.5B	Qwen2.5-0.5B-Instruct	`0.5B_v3`, `0.5B_v4`
1.5B	Qwen2.5-1.5B-Instruct	`1.5B_v1` … `1.5B_v5`
3B	Qwen2.5-3B-Instruct	`3B_v1` … `3B_v5`
7B	Qwen2.5-7B-Instruct	`7B_v1` … `7B_v5`

Each subfolder contains a full model snapshot (config.json, model*.safetensors, tokenizer files, generation_config.json, chat_template.jinja). The vN suffix denotes independent SFT runs (e.g. data/seed/hyperparameter variants).

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer

repo = "Ameame1002/CRAFT"
subfolder = "7B_v1"  # pick any checkpoint from the table above

tok = AutoTokenizer.from_pretrained(repo, subfolder=subfolder)
model = AutoModelForCausalLM.from_pretrained(repo, subfolder=subfolder, device_map="auto")

Or download a single checkpoint:

python

from huggingface_hub import snapshot_download
snapshot_download("Ameame1002/CRAFT", allow_patterns="7B_v1/*", local_dir="./7B_v1")

Architecture (per scale)

All checkpoints use the Qwen2ForCausalLM architecture (32K context):

Scale	hidden	layers	heads	vocab
0.5B	896	24	14	151936
1.5B	1536	28	12	151936
3B	2048	36	16	151936
7B	3584	28	28	152064

License

Released under Apache-2.0, consistent with the upstream Qwen2.5-Instruct models.