CoolFace
Modelpublic

yinita/ps4mas-sft-4b-claude-low400-ep3-mask16k

sourceHugging Faceapache-2.0updated 22d agoView on Hugging Face
0likes335downloads
Model Card

ps4mas-sft-4b-claude-low400-ep3-mask16k

Full-FT Qwen/Qwen3.5-4B reproducing: 9B sft_full_9b_claude_low400_ep3 (tinyeval 3.682); 4B repro with messagelossmask + maxseq=16384

9B originalthis 4B
Datadata/sft_splits/claude_distill/sft_claude_teacher_chat.masked.jsonl (1276 hops)same
Objectivefull sequence SFT (no LoRA)same
Epochs / lr3 / 2e-5same
Batchper-device 1, grad accum 2, 8 GPU ZeRO-2same
Seq4096same

Local path: results/0903/sft_full_4b_claude_low400_ep3_mask16k Uploaded: 2026-09-06T05:33:39.014259+00:00

python
from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained("yinita/ps4mas-sft-4b-claude-low400-ep3-mask16k", trust_remote_code=True)
tok = AutoTokenizer.from_pretrained("yinita/ps4mas-sft-4b-claude-low400-ep3-mask16k", trust_remote_code=True)