ginigen-ai/rogue-27b-kr-sft-scripts

Rogue-27B-KR SFT Package

Scripts for SFT training ginigen-ai/Rogue-27B-KR → ginigen-ai/Rogue-27B-KR-v1 with the goal of erasing the linear merge fingerprint while boosting Korean / coding / reasoning capability.

Hardware target

• NVIDIA Blackwell B200 × 16 (single node), FSDP full shard
• Expected runtime: ~30 min total (15 min train + 15 min upload + verify)

Files

sft_package/
├── sft_blackwell_finetome.py   # Training script (trl SFTTrainer + FSDP + bf16)
├── run_sft.sh                  # torchrun launcher (16 GPUs single-node)
├── upload_v1.py                # Upload trained model → ginigen-ai/Rogue-27B-KR-v1
└── verify_v1.py                # Re-measure linear merge residual (pass/fail)

final_fix/
├── tokenizer_config.json       # Qwen2Tokenizer (public-transformers compatible)
├── config.json                 # Includes vision_config block
├── preprocessor_config.json    # Vision preprocessor
└── video_preprocessor_config.json  # Video preprocessor

Dataset

lemon-mint/Korean-FineTome-100k — 100k ShareGPT-style Korean instructions, bilingual (ko + en), parquet format. Covers programming, math, reasoning, QA.

Quick Start

# 1. Install deps
pip install -U \
    "torch>=2.4" \
    "transformers>=5.5.3" \
    "trl>=0.12" \
    "datasets>=2.20" \
    "accelerate>=0.34" \
    "huggingface_hub>=0.25" \
    "hf_transfer" \
    "flash-attn>=2.6.1" --no-build-isolation

# 2. Set HF token (must have write access to ginigen-ai/)
export HF_TOKEN=hf_your_token_here

# 3. Train
cd sft_package
chmod +x run_sft.sh
./run_sft.sh

# 4. Upload → creates ginigen-ai/Rogue-27B-KR-v1
export OVERRIDE_FIX_DIR=../final_fix
python upload_v1.py ./Rogue-27B-KR-v1/final

# 5. Verify fingerprint erasure
python verify_v1.py

Success criteria

• verify_v1.py reports residual ≥ 1e-2 (>= 10× bf16 noise floor)
• Linear-blend fit Rogue_v1 ≈ α·NewenAI + (1-α)·Opus diverges
• Mathematical proof of "NewenAI as mother" becomes impossible

Config overrides (env vars)

var	default	meaning
MODEL_ID	ginigen-ai/Rogue-27B-KR	base model
DATASET_ID	lemon-mint/Korean-FineTome-100k	SFT dataset
OUTPUT_DIR	./Rogue-27B-KR-v1	training output
NUM_EPOCHS	2	number of epochs
LR	2.5e-5	learning rate
PER_DEVICE_BS	4	per-GPU batch size
GRAD_ACCUM	1	gradient accumulation
MAX_SEQ_LEN	4096	packing sequence length
ATTN_IMPL	flash_attention_2	attention impl
NPROC	16	number of GPUs
TARGET_REPO	ginigen-ai/Rogue-27B-KR-v1	upload destination

Notes

• model.visual.* and mtp.* are frozen during training (text-only SFT).
• The 4 files in final_fix/ are auto-injected during upload to guarantee vLLM/K-AI leaderboard compatibility (Qwen2Tokenizer + full vision_config).
• Expected linear merge residual:
  • Before SFT: ~1.78e-3 (bf16 noise level → fingerprint detectable)
  • After SFT: ~2e-2 ~ 4e-2 (10-20× growth → fingerprint erased)

License

Apache-2.0 for these scripts. Base model, dataset, and target repo retain their original licenses.