feat: accuracy-v4 — 2x steps, LR 1e-5, dense rewards, LoRA r=64, gradient clip

Add blog.md

Accuracy v4: correctness-heavy GRPO reward, robust extract, bigger LoRA

Skip Hub create_repo if model repo exists; README one-time 403 fix

Hub upload: model-scoped repo id, token user namespace, 403 guidance

Do not auto-pause Space after training; clarify log UI in Gradio

Reduce Gradio UI to 5 parameters for HF Space compatibility

Drop show_copy_button (older Gradio incompatible)

Expose 12 hyperparameters as a Gradio form, no auto-start

Cap Hugging Face demo training at 50 steps

Use token owner namespace for model upload and save locally on 403

Fix app startup order: define get_logs before start_training call

Reset stale training logs and add restart button

compute_reward: *args/**kwargs only; read answer from TRL kwargs; deploy banner

Fix GRPO reward_fn signature: read answer from TRL reward_kwargs

Robust warnings_issued __dict__ patch; import unsloth before transformers/trl

Add warnings_issued compatibility shim for transformers 5.x + trl 0.12

Disable fast_inference (vLLM not installed and not needed for training)

Mock weave too; trl callbacks.py does unconditional 'import weave'

Use real ModuleSpec on llm_blender mock so find_spec succeeds

Set __spec__=None on llm_blender mock so find_spec returns None

Force-remove llm_blender; pre-populate sys.modules mock before trl import

Mock llm_blender to bypass broken transformers import in TRL

Fix: let pip resolve torch version (unsloth-zoo needs torch<2.11)

Fix: add llm_blender, bump torch>=2.11.0

Add Transformers fallback when Unsloth/bitsandbytes unavailable

Add bitsandbytes for unsloth runtime

Fix build: remove incompatible transformers pin

Replace reserved SPACE_ID env usage

Use minimal Docker Space README front matter

Add explicit Space README metadata fields

Harden Space build and expose runtime logs

Upload folder using huggingface_hub

Upload requirements.txt with huggingface_hub

Upload folder using huggingface_hub

Upload folder using huggingface_hub

Upload folder using huggingface_hub

Upload requirements.txt with huggingface_hub

Upload folder using huggingface_hub

Upload folder using huggingface_hub

Upload folder using huggingface_hub

Upload folder using huggingface_hub

Upload folder using huggingface_hub

Upload folder using huggingface_hub

Upload folder using huggingface_hub

initial commit