duyentl04/abc

CoT-MTKD Stage 1 experts (Qwen2.5-7B-Instruct, 3 LoRA adapters)

Three LoRA adapters trained jointly in Stage 1 of the CoT_MTKD pipeline. Each adapter is one "expert" on top of Qwen/Qwen2.5-7B-Instruct; the experts are trained with an SFT loss plus a DPP repulsion term that pushes them apart. They are meant to be used as teachers for the Stage 2 distillation step of that pipeline.

Files

Path	Content
expert_0/, expert_1/, expert_2/	PEFT LoRA adapters (adapter_config.json, adapter_model.safetensors)
config.yaml	Stage 1 training config
manifest.json	Run manifest written by the pipeline
metrics.jsonl	Per-step training metrics

The pipeline's own final/adapter_states.pt bundle and the resumable checkpoint.pt referenced in manifest.json are not uploaded. Absolute local paths in config.yaml and manifest.json were rewritten as relative paths, so the config_file_sha256 in the manifest no longer matches config.yaml.

Training setup

• Base model: Qwen/Qwen2.5-7B-Instruct
• LoRA: rank 16, alpha 16, dropout 0.05
• Data: s1K (prepared as s1k_1_1)
• 3 epochs, 94 optimizer steps, global batch size 32, learning rate 5e-5, AdamW, cosine schedule

See config.yaml for the full configuration.

Usage

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = "Qwen/Qwen2.5-7B-Instruct"
tokenizer = AutoTokenizer.from_pretrained(base)
model = AutoModelForCausalLM.from_pretrained(base, torch_dtype="auto", device_map="auto")
model = PeftModel.from_pretrained(model, "duyentl04/abc", subfolder="expert_0")

Replace expert_0 with expert_1 or expert_2 for the other experts.