CoT-MTKD Stage 1 experts (DeepSeek-R1-Distill-Qwen-1.5B, 3 LoRA adapters)

Three LoRA adapters trained jointly in Stage 1 of the CoT_MTKD pipeline (phase1-gac branch, objective sft_dpp_rbf_local_gac). Each adapter is one "expert" on top of deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B; the experts are trained with SFT plus a step-level DPP term, local GAC gradient sharing between experts and RBF repulsion. They are meant to be used as teachers for the Stage 2 distillation step.

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: deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B
• LoRA: rank 16, alpha 16, dropout 0.05, all attention and MLP projections
• Data: s1K-1.1 deepseek_thinking_trajectory, 996 samples
• Prompt: the tokenizer's chat template with the user turn "Please reason step by step, and put your final answer within \boxed{}.\n\n{question}" (the template ends in <think>\n), followed by the CoT and the tokenizer's EOS token
• 3 epochs, 187 optimizer steps, global batch size 16, learning rate 5e-5, AdamW, cosine schedule; first 10% of updates SFT only, then SFT + 0.1 DPP with GAC (beta 0.5) and RBF repulsion (weight 0.5)

See config.yaml for the full configuration.

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base = "deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B"
tokenizer = AutoTokenizer.from_pretrained(base)
model = AutoModelForCausalLM.from_pretrained(base, torch_dtype="bfloat16")
model = PeftModel.from_pretrained(model, "duyentl04/1.5B", subfolder="expert_0")