Fetching metadata from the HF Docker repository...
-
Remove blog_temp.md from repo
-
Remove GitHub mention from blog per HF submission requirements
-
Clean up unused hackathon markdown files and update setup script link
-
Update blog with temp blog content
-
Refine blog.md to improve credibility and tone based on feedback
-
Add final blog post detailing from numeric to semantic reasoning and personal origin
-
README: Update final model repository link
-
README: Update training plots with final winning results (93% compliance, 0->4.5 reward)
-
README: Final results update — 93% risk compliance, 88% governance, semantic reasoning v2.0
-
README: Fix ASCII diagram alignment with standard characters
-
README: Highlight v2.0 Semantic Reasoning (Judge Step 1 Complete)
-
Semantic observation prompts — rich text replaces raw floats (judge feedback #1)
-
Add compulsory Google Colab link to README
-
Update submission blog link to compulsory HF URL
-
Final tweaks to server and requirements
-
Update requirements-space.txt
-
Final manual tweaks to server imports
-
Real ML Demo + Stability Fixes
-
Final Hackathon Polish: Left-side dashboard layout, SVG connection scaling, and 0.4s simulation speedup
-
Hard Forced FP16 precision patch with surgical head casting and stability limits
-
Apply aggressive GRPO stability patch: LR=1e-5, len=64, norm=0.5, BFloat16
-
Update README deliverables with High-Performance training link
-
Disable trainer-level fp16 scaling to resolve GradScaler unscale crash
-
Global Float16 precision patch: align BitsAndBytes, model loading, and GRPOTrainer args
-
Surgical precision fix: explicitly cast lm_head and embed_tokens to compute_dtype
-
Fix precision mismatch during generate() by casting model to compute dtype
-
Patch warnings_issued attribute to fix TRL crash
-
Patch missing TRANSFORMERS_CACHE variable for llm_blender compatibility
-
Refactor train_hf.py to use pure PEFT/Transformers to avoid Unsloth precision bugs
-
Fix LoRA dtype mismatch by explicitly coercing to bfloat16
-
Fix Unsloth chunks crash by setting mini batch to 64
-
Fix Unsloth GRPOConfig compilation attribute name
-
Fix total_memory attribute typo
-
Add standalone HF Jobs GRPO training script (500 steps, 8 gens, sample output logging)
-
Rename WRITEUP.md to blog.md and fix README link
-
Remove broken auto-generated plots in favor of live Kaggle training evidence
-
Update README with requirements alignment, Colab/Kaggle links, and live Kaggle training evidence
-
Delete compiled cache + re-inject Unsloth attrs for Kaggle/Colab compat
-
Bypass Unsloth GRPO compilation - fix SymFloat crash on Colab/Kaggle
-
Inject unsloth_logit_chunk_multiplier into GRPOConfig
-
Inject all expected Unsloth GRPO args
-
Inject unsloth_num_chunks into GRPOConfig
-
Remove unsupported kwargs from GRPOConfig
-
Disable PatchFastRL to fix Colab OSError
-
Fix openenv dependency name in notebook
-
Fix notebook dependencies and CUDA attribute
-
Add Colab GRPO training notebook
-
Add pyarrow to notebook dependencies to fix colab mismatch
-
Fix openenv missing from notebook
-
Update latest changes
