Update model card with pending TB2-lite evaluation status

README.md

@@ -1,55 +1,167 @@
 ---
-base_model: LiquidAI/LFM2.5-8B-A1B
-library_name: peft
+language:
+- en
+- ko
+library_name: transformers
+pipeline_tag: text-generation
 tags:
-- lora
 - terminal
-- tool-use
-- harness-1
-- lfm2
-- lfm2.5
 - sft
-license: other
+- vllm
+- tb2-lite
+- evaluation-pending
+base_model: unknown
 ---
 
-# LFM2.5-8B-A1B Harness-1 LoRA SFT R32 Reloadable
+# LLM-OS-Models/LFM2.5-8B-A1B-Harness1-LoRA-SFT-R32-Reloadable
 
-This repository contains a PEFT/LoRA adapter for `LiquidAI/LFM2.5-8B-A1B` trained on a local Harness-1-style terminal/tool-use SFT bootstrap dataset.
+터미널 작업 자동화를 위한 Terminal SFT 모델입니다. 입력된 작업/이전 터미널 상태를 보고 다음에 실행할 명령을 JSON 형태로 생성하는 용도로 학습했습니다.
 
-## Training Summary
+## 모델 요약
 
-- Base model: `LiquidAI/LFM2.5-8B-A1B`
-- Method: LoRA SFT, rank 32, alpha 64, bf16
-- Context length: 32,768
-- Hardware: 8 x NVIDIA H200
-- Effective batch size: 128
-- Epochs: 3
-- Train rows after packing/cache: 493
-- Runtime: 320.3 seconds
-- Final train loss: 4.673
+- Base model: `unknown`
+- Training setup: `Terminal SFT`
+- Model card snapshot: `2026-06-06 23:10:54 UTC`
+- Corrected TB2-lite evaluated results currently indexed: `60`
+- Corrected TB2-lite score: `pending / not matched in current result directory`
 
-## LoRA Target Modules
+## Quickstart
 
-This adapter intentionally avoids LFM MoE expert `target_parameters` because the current PEFT/Unsloth stack can save those tensors but reload/merge them with mismatched expert dimensions. The reloadable target set is:
+설치와 로그인:
 
-`q_proj,k_proj,v_proj,out_proj,in_proj,w1,w2,w3,gate`
-
-## Data Note
+```bash
+pip install -U vllm transformers huggingface_hub
+huggingface-cli login
+```
 
-The SFT data was generated from public BrowseComp+ style query/gold-document records into deterministic Harness-style tool trajectories. It is a bootstrap run, not the original paper's GPT-5.4 teacher trajectory set.
+관련 코드:
 
-## Use
+- 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`
 
-Load with PEFT on top of the base model:
+vLLM 직접 실행 예시. 평가 코드와 동일하게 chat template을 우선 사용하고, template이 없으면 ChatML/Gemma fallback을 사용합니다.
 
 ```python
-from peft import PeftModel
-from transformers import AutoModelForCausalLM, AutoTokenizer
+from transformers import AutoTokenizer
+from vllm import LLM, SamplingParams
+
+model_id = "LLM-OS-Models/LFM2.5-8B-A1B-Harness1-LoRA-SFT-R32-Reloadable"
+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)
+```
 
-base = "LiquidAI/LFM2.5-8B-A1B"
-adapter = "LLM-OS-Models/LFM2.5-8B-A1B-Harness1-LoRA-SFT-R32-Reloadable"
+권장 출력 형식:
 
-tok = AutoTokenizer.from_pretrained(adapter, trust_remote_code=True)
-model = AutoModelForCausalLM.from_pretrained(base, torch_dtype="auto", trust_remote_code=True)
-model = PeftModel.from_pretrained(model, adapter)
+```json
+{
+  "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
+python tb2_lite/scripts/replay_eval.py \
+  --model LLM-OS-Models/LFM2.5-8B-A1B-Harness1-LoRA-SFT-R32-Reloadable \
+  --model-short LLM-OS-Models__LFM2.5-8B-A1B-Harness1-LoRA-SFT-R32-Reloadable \
+  --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 \
+  --language-model-only
+```
+
+- 기본 권장 tensor parallel: `1`. OOM이면 `--tp`와 `tensor_parallel_size`를 2/4/8로 올리세요.
+- corrected TB2-lite 평가는 `temperature=0.0`, `top_p=1.0`, `max_tokens=1024`로 고정했습니다.
+- Gemma 4는 JSON 출력을 위해 `enable_thinking=False`를 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
+
+- LFM 계열은 빠른 sec/step과 큰 SFT 반응성이 장점입니다. 이 repo는 아직 현재 집계 JSON과 직접 매칭되는 점수가 없어 별도 평가가 필요합니다.
+- 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.