agurung/lcb-med-hard-ft-qwen3-4b-dft-rft-mixed-24-highlr-v1

agurung/lcb-med-hard-ft-qwen3-4b-dft-rft-mixed-24-highlr-v1

PEFT LoRA adapter archive for the LCB med/hard Qwen3-4B FT experiment.

• —Base model: Qwen/Qwen3-4B-Instruct-2507
• —Local run name: dft_rft_mixed_24-lora-r128-a32-lr5e-4-const-lr5e-4-qps8-gpuauto-ep4-highlr_v1
• —Arm: dft_rft_mixed_24
• —Variant: highlr-v1
• —Adapter checkpoints: 14
• —Included bytes: 16027276519
• —Excluded local optimizer/RNG bytes: 0

The archive intentionally keeps HF/PEFT model artifacts and training metadata, but excludes optimizer, scheduler, and RNG state files. These repos are meant for evaluation and RL seeding, not exact optimizer-state resume.

Checkpoint dirs:

checkpoint-10, checkpoint-15, checkpoint-20, checkpoint-25, checkpoint-30, checkpoint-35, checkpoint-40, checkpoint-45, checkpoint-5, checkpoint-50, checkpoint-55, checkpoint-60, checkpoint-65, checkpoint-68