Add README explaining 2026-06-29 cleanup: removed buggy non-strict ckpts

Cleanup 2026-06-29: remove archive_nonstrict/ (buggy non-strict ckpts; global_steps not restored on resume — see strict repo for canonical artifact)

non-strict ckpt-10000 zero_to_fp32.py

non-strict ckpt-10000 vocab.json

non-strict ckpt-10000 training_args.bin

non-strict ckpt-10000 trainer_state.json

non-strict ckpt-10000 tokenizer_config.json

non-strict ckpt-10000 tokenizer.json

non-strict ckpt-10000 special_tokens_map.json

non-strict ckpt-10000 scheduler.pt

non-strict ckpt-10000 rng_state.pth

non-strict ckpt-10000 preprocessor_config.json

non-strict ckpt-10000 non_lora_state_dict.bin

non-strict ckpt-10000 merges.txt

non-strict ckpt-10000 latest

non-strict ckpt-10000 ds-state global_step10000

non-strict ckpt-10000 chat_template.json

non-strict ckpt-10000 added_tokens.json

non-strict ckpt-10000 adapter_model.safetensors

non-strict ckpt-10000 adapter_config.json

non-strict ckpt-9000 zero_to_fp32.py

non-strict ckpt-9000 vocab.json

non-strict ckpt-9000 training_args.bin

non-strict ckpt-9000 trainer_state.json

non-strict ckpt-9000 tokenizer_config.json

non-strict ckpt-9000 tokenizer.json

non-strict ckpt-9000 special_tokens_map.json

non-strict ckpt-9000 scheduler.pt

non-strict ckpt-9000 rng_state.pth

non-strict ckpt-9000 non_lora_state_dict.bin

non-strict ckpt-9000 merges.txt

non-strict ckpt-9000 latest

non-strict ckpt-9000 ds-state global_step9000

non-strict ckpt-9000 chat_template.json

non-strict ckpt-9000 added_tokens.json

non-strict ckpt-9000 adapter_model.safetensors

non-strict ckpt-9000 adapter_config.json

non-strict archive: trainer_state.json

non-strict archive: config.json

non-strict archive: non_lora_state_dict.bin

non-strict archive: adapter_config.json

non-strict archive: adapter_model.safetensors

tensorboard logs

phase2 trainer log

phase2 stdout log

training source (data.py, train.py, ResumeDatasetCallback)

training scripts (sft_phase2.sh, deepspeed config, env)

stage1_merged: Phase1 LoRA merged onto Qwen2.5-VL-7B base (3-expert)

initial commit