Update model card with pending TB2-lite evaluation status

README.md

@@ -1,64 +1,167 @@
-# HRM-Text Ko Terminal B SWE+GLM Pilot
+---
+language:
+- en
+- ko
+library_name: transformers
+pipeline_tag: text-generation
+tags:
+- terminal
+- sft
+- vllm
+- tb2-lite
+- evaluation-pending
+base_model: unknown
+---
 
-Date: 2026-05-23
+# LLM-OS-Models/HRM-Text-Ko-Terminal-B-SWE-GLM-Pilot
 
-This repository contains a raw HRM-Text FSDP2 training checkpoint, not a
-Transformers-ready model. Convert it with `HRM-Text/conversion/convert_to_hf.py`
-after selecting a checkpoint for release.
+터미널 작업 자동화를 위한 Terminal SFT 모델입니다. 입력된 작업/이전 터미널 상태를 보고 다음에 실행할 명령을 JSON 형태로 생성하는 용도로 학습했습니다.
 
-## Run
+## 모델 요약
 
-| Item | Value |
-|---|---:|
-| Architecture | HRM-Text B |
-| Parameters | 435,159,040 |
-| GPUs | 8 x NVIDIA H200 |
-| Epochs | 1 |
-| Global batch | 262,144 tokens |
-| Context | 4,096 train tokens |
-| Wall time | about 7m 38s |
-| Final train loss | 3.00653 |
-| Final token accuracy | 0.46379 |
+- Base model: `unknown`
+- Training setup: `Terminal SFT`
+- Model card snapshot: `2026-05-23 01:03:46 UTC`
+- Corrected TB2-lite evaluated results currently indexed: `56`
+- Corrected TB2-lite score: `pending / not matched in current result directory`
 
-Command:
+## Quickstart
+
+설치와 로그인:
 
 ```bash
-WANDB_MODE=offline WANDB_DIR=/home/work/.data/wandb \
-TOKENIZERS_PARALLELISM=false OMP_NUM_THREADS=1 MKL_NUM_THREADS=1 \
-NCCL_DEBUG=WARN TORCH_NCCL_ASYNC_ERROR_HANDLING=1 \
-torchrun --standalone --nproc_per_node=8 pretrain.py \
-  arch/size@arch=B \
-  data.path=/home/work/.data/hrm_text_prepared/sft_swe_glm_mix_v1 \
-  +checkpoint_path=/home/work/.data/hrm_text_checkpoints/koterm_b_swe_glm_pilot_v1 \
-  +project_name=HRM-Ko-Terminal \
-  +run_name=koterm_b_swe_glm_pilot_v1 \
-  epochs=1 \
-  global_batch_size=262144 \
-  lr_warmup_steps=100 \
-  +log_interval=5 \
-  checkpoint_interval=1
+pip install -U vllm transformers huggingface_hub
+huggingface-cli login
 ```
 
-## Data
+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
+from transformers import AutoTokenizer
+from vllm import LLM, SamplingParams
+
+model_id = "LLM-OS-Models/HRM-Text-Ko-Terminal-B-SWE-GLM-Pilot"
+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)
 
-Prepared dataset:
+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)
 
-`/home/work/.data/hrm_text_prepared/sft_swe_glm_mix_v1`
+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
+{
+  "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/HRM-Text-Ko-Terminal-B-SWE-GLM-Pilot \
+  --model-short LLM-OS-Models__HRM-Text-Ko-Terminal-B-SWE-GLM-Pilot \
+  --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
+```
 
-| Source | Samples | Tokens | Processing |
-|---|---:|---:|---|
-| SWE-ZERO terminal/code trajectories | 53,868 | 182,717,999 | long instructions middle-truncated |
-| GLM-5.1 reasoning cleaned sample | 56,021 | 68,452,781 | `<think>...</think>` stripped, direct answers |
-| Total | 109,889 | 251,170,780 | PrefixLM, response-only loss |
+- 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.
 
-Tokenizer:
+## 평가 상태
 
-`https://huggingface.co/LLM-OS-Models/HRM-Text-Ko-Terminal-Tokenizer-131K`
+- 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.
 
-## Files
+## 모델군 해석
 
-- `fsdp2_epoch_1/`: distributed FSDP2 model and optimizer checkpoint.
-- `carry_epoch_1.*.pt`: per-rank carry state.
-- `all_config.yaml`: resolved training config.
-- `train_metadata.yaml`: tokenizer and dataset metadata.
-- `hrm_nocarry_bp_warmup.py`: copied model source for reproducibility.
+- 이 repo는 아직 현재 corrected TB2-lite 집계 JSON과 직접 매칭되는 점수가 없습니다.
+- 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.