-
Add TurboQuant 3-bit + HQQ mixed-precision compression pipeline (sections 9-11)
-
Rewrite from scratch — exact copy of official unsloth Qwen2.5-7B notebook pattern with pinned deps
-
Speed fix v3: seq 1024, bs 4, max_steps=800, max_length param, LoRA r=16 — target ~2h on T4
-
Speed fix: packing=True, seq 2048, bs 2, 1 epoch — estimated ~2h on T4 instead of 177h
-
Fix: train_on_responses_only needs explicit Qwen2.5 ChatML parts + warmup_ratio→warmup_steps + total_mem→total_memory
-
Upload generate_v2_data.py with huggingface_hub
-
Upload WebScrapeAgent_Training_v2.ipynb with huggingface_hub
-
Upload WebScrapeAgent_Training.ipynb with huggingface_hub
-
Upload evaluate.py with huggingface_hub
-
Upload webscrape_agent.py with huggingface_hub
-
Upload train.py with huggingface_hub
-
Upload README.md with huggingface_hub
-
initial commit
