You can deploy this model and start earning money today!

Gabriel2502/
DeepSeek-R1-Distill-Qwen-7B-latent-sft-stage2_v2

Name: DeepSeek-R1-Distill-Qwen-7B-latent-sft-stage2_v2
Availability: InStock

deepseek-ai/DeepSeek-R1-Distill-Qwen-7B

adapter

llama3

Views

No views yet

LatentSFT Stage 2 — DeepSeek-R1-Distill-Qwen-7B-latent-sft-stage2_v2

Model ini adalah hasil fine-tuning deepseek-ai/DeepSeek-R1-Distill-Qwen-7B menggunakan metode Latent SFT Stage 2 dengan LoRA (Low-Rank Adaptation).

Konfigurasi Training

Parameter	Nilai
Base Model	`deepseek-ai/DeepSeek-R1-Distill-Qwen-7B`
LoRA Rank	32
LoRA Dropout	0.1
Compression Rate	4
Epochs	20
Batch Size	1
Gradient Accumulation	4
Learning Rate	0.0001
Max Seq Len	1024
CE Weight	1.0
KL Weight	0.5
FP16	True

Cara Load Model

python

from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel

base_model = AutoModelForCausalLM.from_pretrained("deepseek-ai/DeepSeek-R1-Distill-Qwen-7B")
tokenizer  = AutoTokenizer.from_pretrained("Gabriel2502/DeepSeek-R1-Distill-Qwen-7B-latent-sft-stage2_v2")
model      = PeftModel.from_pretrained(base_model, "Gabriel2502/DeepSeek-R1-Distill-Qwen-7B-latent-sft-stage2_v2")
model      = model.merge_and_unload()  # Opsional: merge LoRA ke base model

Deskripsi Metode

Model ini dilatih menggunakan pendekatan Latent Compression:

Jawaban (response) dikompresi menjadi latent token menggunakan rata-rata distribusi token
Model belajar memprediksi jawaban dari representasi latent yang terkompresi
Loss gabungan: CE Loss (pada token jawaban) + KL Loss (distilasi distribusi latent)