fufugita/blackhole-mix-v2

Add GGUF conversion job (F16 then Q4_K_M via llama.cpp)

Fix: use quantization_config (not load_in_4bit kwarg)

Add merge job script

Fix: chat template + max_tokens=2048 + enable_thinking=False

Use bundled labs (no git clone)

Bundle black-hole eval labs for offline job use

Add eval job script for LoRA adapter vs black-hole labs

T4 OOM fix: batch=1, grad_accum=8, grad_checkpointing, expandable_segments

Self-install deps for hf jobs run form

Add distill-base training script

Fix input construction for generate (return_dict + mask)

Use sdpa attention (no flash-attn in job env)

Drop kernel deps; test fallback path

Rewrite probe as uv script with inline deps

Add distill probe script

Upload train_job.py with huggingface_hub

Upload train_job.py with huggingface_hub

Upload train_job.py with huggingface_hub

Upload train_job.py with huggingface_hub

Upload train_job.py with huggingface_hub

Upload data/train.jsonl with huggingface_hub

initial commit