FactGuard-9B-LoRA

🇰🇷 Korean card: MODEL_CARD.ko.md

These are the LoRA adapter weights that were merged to build ThakiCloud/FactGuard-9B. They are published so the training can be reproduced, and so the model can be loaded as an adapter instead of as merged weights. Intended use, training data, evaluation and limitations are documented on the merged model's card. This repo does not change any of them.

Stages

#	stage	where	trained on	r / alpha
1	v0-sft	v0-sft/	earlier checkpoint of the same adapter; the final stage continued training it, so only the final stage is applied	16 / 32
2	v0.1-sft	v0.1-sft/	earlier checkpoint of the same adapter; the final stage continued training it, so only the final stage is applied	16 / 32
3	v0.3-sft	v0.3-sft/	Qwen/Qwen3.5-9B	16 / 32

Only the stages used in the code below are applied; the others are earlier checkpoints kept for the record.

How to use

import torch
from transformers import AutoModelForCausalLM
from peft import PeftModel

model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3.5-9B", torch_dtype=torch.bfloat16)
model = PeftModel.from_pretrained(model, "ThakiCloud/FactGuard-9B-LoRA", subfolder="v0.3-sft").merge_and_unload()  # stage 3: v0.3-sft

How we checked these are the right adapters

For 4 randomly chosen LoRA target tensors we compared the published merged weights minus Qwen/Qwen3.5-9B with the adapter deltas (scale x B x A). The median cosine is 0.978. Rounding the merged weights to bf16 keeps it below 1; an adapter that was not used scores about 0.

Notes

• adapter_config.json names the public model each stage was trained on. Training used a local copy of the same weights, and the adapter weights are unchanged.
• Earlier stages are kept for reproducibility only. Each stage continued training the same LoRA weights, so load only the final stage on top of the base model (as in the code above).

License

Same as ThakiCloud/FactGuard-9B: apache-2.0.