It is trained for Korean legal retrieval tool use, not for legal answer generation. The model reads a legal query and emits JSON tool actions such as fan_out_search, review_docs, curate, verify, and end_search.

Intended Use

Retrieve Korean statutes, precedents, administrative rules, and ordinances.
Select evidence doc_id values for a downstream LLM.
Package selected evidence into RAG context.

Do not use this adapter as a standalone legal advice model.

Local Evaluation

Source-rag1200 post-hoc/regression evaluation:

metric	value
count	1200
recall	0.67917
precision	0.53042
F2	0.58451
trajectory_recall	0.98583
valid_action_rate	0.82119
ended	0.56167

This is not a clean hidden benchmark because the source-rag data was used in repair training. Treat it as regression/post-hoc repair evidence.

vLLM Example

Install:

pip install -U "huggingface_hub[cli]" peft transformers vllm

Download the LoRA adapter:

bash

hf download gyung/lfm25-ko-legal-retriever-v10-source-repair-lora \
  --local-dir lfm25-ko-legal-retriever-v10-source-repair-lora

Serve the base model with this adapter:

bash

python -m vllm.entrypoints.openai.api_server \
  --model LiquidAI/LFM2.5-8B-A1B \
  --enable-lora \
  --lora-modules lfm25-ko-legal-retriever-v10=./lfm25-ko-legal-retriever-v10-source-repair-lora \
  --served-model-name lfm25-ko-legal-retriever-v10 \
  --max-model-len 12288 \
  --gpu-memory-utilization 0.92

Request:

bash

curl http://127.0.0.1:8000/v1/chat/completions \
  -H 'Content-Type: application/json' \
  -d '{
    "model": "lfm25-ko-legal-retriever-v10",
    "messages": [
      {"role": "system", "content": "assistant 응답은 항상 JSON action 하나만 출력합니다. 허용 tool: fan_out_search, review_docs, curate, verify, end_search."},
      {"role": "user", "content": "민법 제126조 표현대리와 관련 판례 근거 문서를 찾아라."}
    ],
    "temperature": 0.1,
    "max_tokens": 192
  }'

Expected style:

{"tool":"fan_out_search","queries":["민법 제126조 표현대리","표현대리 정당한 이유 판례","무권대리 표현대리 대법원"]}

PEFT Loading Example

python

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base_id = "LiquidAI/LFM2.5-8B-A1B"
adapter_id = "gyung/lfm25-ko-legal-retriever-v10-source-repair-lora"

tokenizer = AutoTokenizer.from_pretrained(base_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    base_id,
    device_map="auto",
    torch_dtype="auto",
    trust_remote_code=True,
)
model = PeftModel.from_pretrained(model, adapter_id)

messages = [
    {
        "role": "system",
        "content": "assistant 응답은 항상 JSON action 하나만 출력합니다. 허용 tool: fan_out_search, review_docs, curate, verify, end_search.",
    },
    {"role": "user", "content": "민법 제126조 표현대리와 관련 판례 근거 문서를 찾아라."},
]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=192, do_sample=False)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True))

Artifacts

Local training/eval artifacts are indexed in harness-1/README.md and harness-1/artifacts/README.md.