gemma-4-E4B-it-Terminal-SFT-2Epoch-DDP-4GPU

모델 요약

• Base model: google/gemma-4-E4B-it
• Training setup: 2 epochs, DDP fine-tuning
• Model card snapshot: 2026-06-03 22:09:28 UTC
• Corrected TB2-lite evaluated results currently indexed: 60
• Corrected TB2-lite score: pending / not matched in current result directory

Quickstart

설치와 로그인:

bash
Copy code
pip install -U vllm transformers huggingface_hub
huggingface-cli login

관련 코드:

• GitHub: https://github.com/LLM-OS-Models/Terminal
• vLLM 평가 실행: tb2_lite/scripts/replay_eval.py
• chat template/fallback 생성: tb2_lite/scripts/prompt_builder.py
• JSON/command 채점: tb2_lite/scripts/replay_metrics.py

vLLM 직접 실행 예시. 평가 코드와 동일하게 chat template을 우선 사용하고, template이 없으면 ChatML/Gemma fallback을 사용합니다.

python
Copy code
from transformers import AutoTokenizer
from vllm import LLM, SamplingParams

model_id = "LLM-OS-Models/gemma-4-E4B-it-Terminal-SFT-2Epoch-DDP-4GPU"
tp = 1

tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
llm = LLM(
    model=model_id,
    tokenizer=model_id,
    trust_remote_code=True,
    dtype="bfloat16",
    tensor_parallel_size=tp,
    max_model_len=49152,
    gpu_memory_utilization=0.92,
)

messages = [
    {"role": "system", "content": "You are a terminal automation assistant. Return JSON only."},
    {"role": "user", "content": "Inspect the current directory and list Python files."},
]

def render_chatml(messages):
    parts = []
    for message in messages:
        role = "assistant" if message["role"] == "assistant" else message["role"]
        if role == "tool":
            role = "user"
        parts.append(f"<|im_start|>{role}\n{message['content']}<|im_end|>\n")
    parts.append("<|im_start|>assistant\n")
    return "".join(parts)

def render_gemma4_turn(messages, empty_thought_channel=False):
    parts = ["<bos>"]
    for message in messages:
        role = "model" if message["role"] == "assistant" else message["role"]
        if role == "tool":
            role = "user"
        parts.append(f"<|turn>{role}\n{message['content'].strip()}<turn|>\n")
    parts.append("<|turn>model\n")
    if empty_thought_channel:
        parts.append("<|channel>thought\n<channel|>")
    return "".join(parts)

def render_prompt(model_id, tokenizer, messages):
    model_key = model_id.lower()
    if "gemma-4" in model_key:
        try:
            return tokenizer.apply_chat_template(
                messages,
                tokenize=False,
                add_generation_prompt=True,
                enable_thinking=False,
            )
        except Exception:
            return render_gemma4_turn(
                messages,
                empty_thought_channel=("26b" in model_key or "31b" in model_key),
            )
    try:
        return tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
    except Exception:
        return render_chatml(messages)

prompt = render_prompt(model_id, tokenizer, messages)
sampling = SamplingParams(
    temperature=0.0,
    top_p=1.0,
    max_tokens=1024,
    repetition_penalty=1.0,
)
outputs = llm.generate([prompt], sampling_params=sampling)
print(outputs[0].outputs[0].text)

권장 출력 형식:

json
Copy code
{
  "analysis": "brief reasoning about the next terminal action",
  "plan": "short execution plan",
  "commands": [
    {"keystrokes": "ls -la\n", "duration": 0.1}
  ],
  "task_complete": false
}

평가와 동일한 replay 명령:

bash
Copy code
python tb2_lite/scripts/replay_eval.py \
  --model LLM-OS-Models/gemma-4-E4B-it-Terminal-SFT-2Epoch-DDP-4GPU \
  --model-short LLM-OS-Models__gemma-4-E4B-it-Terminal-SFT-2Epoch-DDP-4GPU \
  --eval-path tb2_lite/data/replay_full.jsonl \
  --output-dir /home/work/.data/tb2_lite_eval/corrected_readme_models_vllm \
  --dtype bfloat16 \
  --tp 1 \
  --max-model-len 49152 \
  --max-tokens 1024 \
  --temperature 0.0 \
  --top-p 1.0 \
  --gpu-memory-utilization 0.92 \
  --thinking-mode off \
  --strip-thinking-history auto \
  --gemma4-empty-thought-channel auto \
  --language-model-only

• 기본 권장 tensor parallel: 1. OOM이면 --tp와 tensor_parallel_size를 2/4/8로 올리세요.
• corrected TB2-lite 평가는 temperature=0.0, top_p=1.0, max_tokens=1024로 고정했습니다.
• Gemma 4는 JSON 출력을 위해 enable_thinking=False를 사용하고, 26B/31B 계열은 평가 코드에서 empty thought channel 처리를 자동 적용합니다.

평가 상태

• Current corrected TB2-lite score: pending
• Reason: 현재 /home/work/.data/tb2_lite_eval/corrected_readme_models_vllm 집계 결과와 이 HF repo명이 직접 매칭되지 않았습니다.
• Next step: 동일한 tb2_lite/scripts/replay_eval.py 경로로 평가를 돌린 뒤 점수 카드로 자동 교체합니다.

모델군 해석

• Gemma 계열은 native Gemma/Liquid 전처리와 chat template 처리가 중요합니다. 이 repo는 corrected 평가가 끝나면 점수 카드로 교체합니다.
• TB2-lite 점수는 일반 지능 벤치마크가 아니라 터미널 next-action JSON 재현 능력을 측정합니다.
• 생성 명령은 실제 실행 전에 sandbox, allowlist, human review 같은 안전장치를 거쳐야 합니다.