Downloads · 30 days
20
6% of all-time downloads
haofue2i1z3/Qwen3-8B-LegalCitation-GRPO-600
Qwen3-8B-LegalCitation-GRPO-600 is a text generation model from haofue2i1z3. Use it when you need the model to write or continue text. It is set up for peft. The card lists the license as apache-2.0.
从 GRPO-200 checkpoint 按原配方继续训练至全局 600 步的剂量响应终点。该模型用于检验残余问题是否仅由训练剂量不足导致。
Downloads · 30 days
20
6% of all-time downloads
All-time downloads
320
Public
Repo size
710 MB
Likes
0
Public
Click a slice to open those files.
.safetensors698 MB · 98%
From the Hugging Face model README
从 GRPO-200 checkpoint 按原配方继续训练至全局 600 步的剂量响应终点。该模型用于检验残余问题是否仅由训练剂量不足导致。
dev84 的 required 行召回率由 Base 的 83.33% 变为 88.10%,低于 200 步的 92.86% 高水位;grounded citation precision 为 77.36%,仍远低于 95% 门禁。Legal-50 统一场均分为 Base 6.908、200 步 6.924、600 步 6.944,差异远低于约 0.4 的可辨别尺度。
最终判定为 FAIL。600 步没有证明额外剂量能够解决剩余引用、停止和重复问题,因此该模型只作为剂量响应研究终点发布。
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
adapter = "haofue2i1z3/Qwen3-8B-LegalCitation-GRPO-600"
base = "Qwen/Qwen3-8B"
tokenizer = AutoTokenizer.from_pretrained(adapter, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(base, device_map="auto", torch_dtype="auto")
model = PeftModel.from_pretrained(model, adapter)