README.md

---
base_model: llm-jp/llm-jp-3-13b
tags:
- text-generation-inference
- transformers
- unsloth
- llama
- trl
license: apache-2.0
language:
- en
- ja
datasets:
- weblab-GENIAC/aya-ja-evol-instruct-calm3-dpo-masked
- DeL-TaiseiOzaki/Tengentoppa-sft-mini-vol1.0
---

## 以下は推論用コードです。

* 事前に以下をインストールしてください。
  * pip install -q numpy==1.26.4
  * pip install -q vllm==0.6.4
  * pip install -q bitsandbytes==0.44.1



```python

from vllm import LLM, SamplingParams
from vllm.lora.request import LoRARequest

import torch
import json

from datasets import load_dataset
from huggingface_hub import snapshot_download

id = "llm-jp-3-13b-it-bs4-ac10-step370-lora"
lora_path = snapshot_download(repo_id="jaked97/"+ id)
model_id = "models/models--llm-jp--llm-jp-3-13b/snapshots/cd3823f4c1fcbb0ad2e2af46036ab1b0ca13192a"

tasks = load_dataset("json", data_files="./elyza-tasks-100-TV_0.jsonl", split="train")


prompts = [
    f"""### instruction:
あなたは親切なAIアシスタントです。
### input:
{input}
### output:
""" for input in tasks["input"]]

llm = LLM(
    model=model_id,
    gpu_memory_utilization=0.99,
    quantization="bitsandbytes",
    load_format="bitsandbytes",
    trust_remote_code=True,
    enforce_eager=True,
    enable_lora=True,
    max_lora_rank=64,
)

outputs = llm.generate(
    prompts,
    sampling_params = SamplingParams(
        temperature=0,
        max_tokens=1024,
        min_tokens=1,
        repetition_penalty=1.2,
        skip_special_tokens=True,
        seed=97,
    ),
    lora_request=LoRARequest("sql_adapter", 1, lora_path),
)

with open(f"./{id}_max1024-nf4-vllm.jsonl", 'w', encoding='utf-8') as f:
    for i in range(len(outputs)):
        result = {
            "task_id" : tasks[i]["task_id"],
            "input" : tasks[i]["input"],
            "output" : outputs[i].outputs[0].text
        }
        json.dump(result, f, ensure_ascii=False)
        f.write('\n')

```

# Uploaded  model

- **Developed by:** jaked97
- **License:** apache-2.0
- **Finetuned from model :** llm-jp/llm-jp-3-13b

This llama model was trained 2x faster with [Unsloth](https://github.com/unslothai/unsloth) and Huggingface's TRL library.

[<img src="https://raw.githubusercontent.com/unslothai/unsloth/main/images/unsloth%20made%20with%20love.png" width="200"/>](https://github.com/unslothai/unsloth)