---
language:
- ko
- en
library_name: pytorch
pipeline_tag: text-generation
base_model: LLM-OS-Models/KoHRM-Text-1.4B
base_model_relation: finetune
tags:
- kohrm
- hrm-text
- terminal-agent
- tool-use
- full-sft
- prefixlm
---

# KoHRM-Text-1.4B FullSFT Top2 Terminal Tool Merge Epoch1

This is an experimental full-SFT checkpoint for the KoHRM-Text 1.4B PrefixLM runtime.

## Base Model

- Base model: `LLM-OS-Models/KoHRM-Text-1.4B`
- Relation: full fine-tune (`base_model_relation: finetune`)
- This repository contains full fine-tuned KoHRM-Text weights exported as `model.safetensors`.

It is a fine-tuned version of `LLM-OS-Models/KoHRM-Text-1.4B`. The training resumed from the stage4d KoHRM checkpoint with a merged terminal/tool dataset built from the current top LFM2.5 terminal SFT runs. The goal is to move KoHRM from generic PrefixLM generation toward TB2-lite terminal next-action JSON outputs.

## Training

- Base model: `LLM-OS-Models/KoHRM-Text-1.4B`
- Dataset: `kohrm_sft_top2_terminal_tool_raw8192_v1`
- Context length: `8192`
- Approximate training tokens: `245M`
- Training type: full SFT, not LoRA
- Epochs: `1`
- GPUs: `4 x H200`
- Global batch size: `90112` tokens
- Learning rate: `2e-5`
- Export format: single-file `model.safetensors` plus tokenizer/config files

## Evaluation

TB2-lite full replay evaluation completed on 2026-06-05 KST.

| Checkpoint | Steps | Score | Cmd F1 | Precision | Recall | First Cmd | Valid JSON |
| --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: |
| `303/303 full replay` | 303 | 31.59 | 0.3159 | 0.3859 | 0.3415 | 24.8% | 73.3% |

This final score is above the best completed KoHRM LoRA result (`29.11`, +2.48) and just below `LLM-OS-Models/Ouro-1.4B-Thinking-Terminal-SFT` (`31.74`, -0.15). The main gain over the LoRA runs is better command recall and JSON stability after moving the base weights directly with full SFT. The remaining gap to the stronger Qwen/LFM terminal SFT models is mostly command coverage and first-action accuracy.

The previous KoHRM stage4d results before this full SFT were:

| Model | Score | Cmd F1 | Precision | Recall | First Cmd | Valid JSON |
| --- | ---: | ---: | ---: | ---: | ---: | ---: |
| `KoHRM-Text-1.4B-stage4d direct` | 11.48 | 0.1148 | 0.1995 | 0.0961 | 5.9% | 38.9% |
| `stage4d + terminal-tool-core-r64 LoRA` | 29.11 | 0.2911 | 0.3988 | 0.2768 | 22.1% | 63.4% |
| `LLM-OS-Models/Ouro-1.4B-Thinking-Terminal-SFT` | 31.74 | 0.3174 | 0.4062 | 0.3410 | 24.8% | 63.7% |

This full-SFT checkpoint is the best completed KoHRM full-weight result in this repository at upload time, but it should still be treated as experimental because the local HRM-Text PrefixLM runtime is slower than the vLLM chat-model path used by most leaderboard entries.

## Usage Notes

This is not a standard Hugging Face `AutoModelForCausalLM` export yet. It is a KoHRM/HRM-Text PrefixLM checkpoint and currently requires the local HRM-Text runtime.

The local evaluation path used for this export is:

```bash
python tb2_lite/scripts/replay_eval_hrm_text.py \
  --model /path/to/KoHRM-Text-1.4B-fullsft-top2-terminal-tool-merge-epoch1 \
  --model-short KoHRM-Text-1.4B-fullsft-top2-terminal-tool-merge-epoch1 \
  --eval-path tb2_lite/data/replay_full.jsonl \
  --output-dir tb2_lite/results/kohrm_fullsft_top2_export \
  --local-hrm-export \
  --base-ckpt-path /path/to/KoHRM-Text-1.4B-fullsft-top2-terminal-tool-merge-gbs180k-4gpu \
  --max-model-len 4096 \
  --max-tokens 1024 \
  --batch-size 16
```

For now, use the repository README benchmark table as the source of truth for completed scores.