Downloads · 30 days
18
8% of all-time downloads
gyung/lfm2-1.2b-koen-mt-v6.4-merged
lfm2-1.2b-koen-mt-v6.4-merged is a text generation model from gyung. Use it when you need the model to write or continue text. The card lists the license as other.
LiquidAI LFM2-1.2B 기반 고품질 한영/영한 양방향 번역 모델 (SFT)
Downloads · 30 days
18
8% of all-time downloads
All-time downloads
223
Public
Parameters
1.2B
33.1 GB on disk
Likes
1
Public
Click a slice to open those files.
.pt9.5 GB · 58%
From the Hugging Face model README
LiquidAI LFM2-1.2B 기반 고품질 한영/영한 양방향 번역 모델 (SFT)
🎯 v8 RL Adapter의 Base Model입니다.
RL Adapter: gyung/lfm2-1.2b-koen-mt-v8-rl-10k-adapter (CHrF++ 34.61 🏆)
| Rank | Model | CHrF++ | BLEU | Params |
|---|---|---|---|---|
| 1 | Google Translate | 39.27 | 18.18 | - (API) |
| 2 | LFM2-v8-RL (with Adapter) | 34.61 | 13.21 | 1.2B |
| 3 | LFM2-v6.4-merged (This) | 33.53 | 12.63 | 1.2B |
| 4 | Gemma-3-4B-it-GGUF | 32.83 | 11.36 | 4B |
| 5 | LFM2-1.2B (Base) | 27.23 | 6.43 | 1.2B |
✅ v6.4는 GRPO 학습에 최적화된 안정적인 베이스 모델
"" 패턴이 완전히 사라짐💡 권장: 프로덕션 환경에서는 RL Adapter가 적용된 v8 사용 권장
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
model_id = "gyung/lfm2-1.2b-koen-mt-v6.4-merged"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
model_id,
device_map="auto",
torch_dtype=torch.float16
)
messages = [
{"role": "system", "content": "Translate to Korean."},
{"role": "user", "content": "The training is progressing smoothly."}
]
input_ids = tokenizer.apply_chat_template(
messages,
return_tensors="pt",
add_generation_prompt=True
).to(model.device)
outputs = model.generate(
input_ids,
max_new_tokens=256,
do_sample=True,
temperature=0.3,
min_p=0.15,
repetition_penalty=1.05
)
decoded = tokenizer.decode(outputs[0][input_ids.shape[1]:], skip_special_tokens=True)
print(decoded)
# 출력: 학습이 순조롭게 진행되고 있습니다.
messages = [
{"role": "system", "content": "Translate to English."},
{"role": "user", "content": "오늘 날씨가 정말 좋습니다."}
]
# 동일한 generate 코드 사용
# 출력: The weather is really nice today.
| Phase | Dataset | Samples | Focus |
|---|---|---|---|
| Phase 1 | v6-200k | 200,000 | 일반 도메인 + 관용구 |
| Phase 2 | v6.1-80k (Filtered) | 80,611 | CometKiwi > 0.85 고품질 |
| 항목 | 값 |
|---|---|
| Base Model | LiquidAI/LFM2-1.2B |
| Method | Full Fine-tuning (SFT) |
| Total Steps | ~4,000 |
| Final Loss | 1.2429 |
| Token Accuracy | 71.60% |
| Hardware | Kaggle T4 x 2 |
| Optimizer | paged_adamw_8bit |
| Learning Rate | 5e-6 (Cosine Decay) |
| Step | Loss | Token Acc | 비고 |
|---|---|---|---|
| 750 | 1.28 | - | - |
| 1500 | 1.2578 | 71.27% | Healthy |
| 3000 | 1.2506 | 71.47% | Stable |
| 4000 | 1.2429 | 71.60% | Final |
@misc{lfm2-koen-v6.4,
author = {gyung},
title = {LFM2-1.2B-KoEn-MT-v6.4: SFT-Enhanced Bidirectional Korean-English Translation Model},
year = {2024},
publisher = {Hugging Face},
url = {https://huggingface.co/gyung/lfm2-1.2b-koen-mt-v6.4-merged}
}
이 모델은 Liquid AI LFM Open License v1.0을 따릅니다.