EVAFRILL-Mo-3B / sft-v2 /README.md
pathcosmos's picture
Upload sft-v2/README.md with huggingface_hub
b229e78 verified
|
Raw
History Blame Contribute Delete
1.44 kB
metadata
language:
  - ko
  - en
license: apache-2.0
tags:
  - sft
  - instruction-tuned
  - chat
  - korean
  - llm
pipeline_tag: text-generation

EVAFRILL-Mo 3B — SFT v2

Instruction-tuned variant of EVAFRILL-Mo 3B. Fine-tuned on Korean/English instruction data with NEFTune noise augmentation.

Training Stage

Supervised Fine-Tuning (SFT) on top of the pretrained base checkpoint.

Key Details

  • Steps: 65,000 (early stopped)
  • Stop criterion: Validation loss plateau at 1.79
  • NEFTune alpha: 5.0
  • Gradient Checkpointing: enabled
  • Precision: BF16

Metrics

Metric Value
Validation loss (final) 1.79

Chat Template

<|user|>
{user message}
<|assistant|>
{assistant response}

Notes

This is the primary instruction-following checkpoint. It serves as the base for DPO rounds and the SLERP merge. For best results with reduced repetition, consider using the SLERP variant instead.

Main Model Card

See the main README for full project details, architecture, and training history.

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained("path/to/sft-v2", torch_dtype="bfloat16")
tokenizer = AutoTokenizer.from_pretrained("path/to/sft-v2")
inputs = tokenizer("<|user|>\n질문을 여기에 입력하세요\n<|assistant|>\n", return_tensors="pt")