Ameame1002/CRAFT

Does Faithfulness-Guided Alignment Hurt Accuracy?

Unlocking Accurate and Faithful Post-Retrieval Reasoning

Structured traces · Rule-based rewards · Semantic faithfulness audits

CRAFT GRPO checkpoints based on the Qwen2.5-Instruct family.

Accepted to the EMNLP 2026 Main Conference. Read the paper.

Repository Layout

Model scale	Base model	Available directories
0.5B	Qwen2.5-0.5B-Instruct	0.5B_v1, 0.5B_v2, 0.5B_v3, 0.5B_v4, 0.5B_v5
1.5B	Qwen2.5-1.5B-Instruct	1.5B_v1, 1.5B_v2, 1.5B_v3, 1.5B_v4, 1.5B_v5
3B	Qwen2.5-3B-Instruct	3B_v1, 3B_v2, 3B_v3, 3B_v4, 3B_v5
7B	Qwen2.5-7B-Instruct	7B_v1, 7B_v2, 7B_v3, 7B_v4, 7B_v5

Every directory follows the same inference-ready layout:

<scale>_<variant>/
  config.json
  generation_config.json
  model.safetensors                 # or model-*.safetensors with an index
  tokenizer.json
  tokenizer_config.json
  chat_template.jinja
  added_tokens.json
  special_tokens_map.json
  merges.txt
  vocab.json

Weight-shard counts vary with model size. Sharded models also include model.safetensors.index.json. Training logs, optimizer states, and trainer state files are not included.

Checkpoint Provenance

The 0.5B, 1.5B, and 3B v1-v4 models are local 312-step GRPO checkpoints. The 7B models were restored from the archived Hub upload; they are not the new local 312-step model series. The small-model v5 directories retain the legacy GRPO_* weights unchanged; they are not the newer local 312-step v5 checkpoints. SFT checkpoints are not part of this release.

Manuscript-reported scores include fitted simulations and counterfactual estimates; they should not be treated as measured results of these downloads. See results provenance.

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

repo = "Ameame1002/CRAFT"
subfolder = "7B_v1"

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

Citation

@misc{liu2026craft,
  title = {Does Faithfulness-Guided Alignment Hurt Accuracy? Unlocking Accurate and Faithful Post-Retrieval Reasoning},
  author = {Liu, Yu and Zhang, Wenxiao and Guo, Diandian and Cao, Cong and Yuan, Fangfang and Sun, Qiang and Liu, Yanbing and Hong, Jin B. and Ma, Zhiyuan},
  year = {2026},
  eprint = {2602.01348},
  archivePrefix = {arXiv},
  primaryClass = {cs.CL},
  doi = {10.48550/arXiv.2602.01348},
  url = {https://arxiv.org/abs/2602.01348},
  note = {Accepted to EMNLP 2026 Main Conference}
}

License

Apache-2.0.