Modelpublic

yinita/ps4mas-sft-mt-glm52-9b-v2

sourceHugging Faceapache-2.0updated 23d agoView on Hugging Face

0likes24downloads

Model Card

ps4mas-sft-mt-glm52-9b-v2

LoRA SFT on Qwen/Qwen3.5-9B with strict message_loss_mask over glm52 multitrack v2: sft2_acquire (importance top-3 keys only, empty-content acquire_user_info) + sft3_final_balanced (composite≥3, upsampled), history tool names canonicalized (acquire_user_info / query_memory).


Data	`data/sft_splits/glm52_multitrack/sft_masked_train.jsonl` (7232 rows)
Method	LoRA r=16 α=32, masked SFT
Epochs / lr	3 / 1e-4
Seq	4096

Local path: results/0903/sft_mt_glm52_9b_v2 Uploaded: 2026-09-04T17:13:32.115358+00:00

python

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer
base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3.5-9B", trust_remote_code=True)
model = PeftModel.from_pretrained(base, "yinita/ps4mas-sft-mt-glm52-9b-v2")
tok = AutoTokenizer.from_pretrained("yinita/ps4mas-sft-mt-glm52-9b-v2", trust_remote_code=True)