Exclude large jsonl files from repo

Add git lfs tracking for large jsonl files

Fix Gradio/huggingface_hub version compatibility

Add ddgs for free web search

Add real web search with ddgs (DuckDuckGo HTML scraper, free, no API key)

Fix search with DuckDuckGo API (proper JSON parsing)

Use local model instead of HF for faster loading

Add web search command (search:query) with DuckDuckGo API

Load model from HF: my-ai-stack/stack-2-9-finetuned

Improve chat.py with system prompt and User/Assistant format

Add interactive chat script with improved generation settings

feat: add code completion generator and model registry tools

feat: add 5000 more tool calling examples (total: 6500)

feat: add training recipes for T4 QLoRA, A100, local training

feat: add evaluation datasets (HumanEval 50, MBPP 100, Tool scenarios 50)

feat: add production infrastructure - CI/CD, Docker, code quality, and monitoring

feat: add inference API, quickstart guide, roadmap, and combined tool data

feat: add evaluation scripts, tool calling data generator, and 7B training configs

fix: remove Google Drive mount (Colab-only) - use Kaggle output + GitHub push instead

feat: add Google Drive auto-save to prevent losing model outputs

feat: add model testing and evaluation scripts

fix: add numpy import at top to ensure it's loaded (fixes Numpy not available error)

fix: disable AMP (fp16=False, bf16=False) to bypass P100 GradScaler bug

fix: load model in FP32 to avoid AMP gradient scaling conflict

fix: proper variable ordering for bf16/fp16 detection — define before use

fixes for Kaggle run

fix: define use_fp16 before model load; remove duplicate

fix: load model in fp16 to match fp16 training precision (bf16 not supported on P100/T4)

fix: force FP16 and disable BF16 for T4 compatibility

fix: use torch.cuda.is_bf16_supported() instead of compute capability check

fix: load model in bfloat16, train in fp16 on T4 (bf16 not supported on Turing)

fix: set expandable_segments=False to fix PyTorch #124807 gradient checkpointing bug

fix: use device_map={'': cuda} to force all model layers on GPU (fixes gradient flow issue)

fix: use device_map=auto for 1.5B model, add model.train() after LoRA

chore: use Qwen2.5-Coder-1.5B for Kaggle T4 (7B doesn't fit in 16GB)

fix: use device_map=None for single GPU (T4) to avoid meta tensor errors with float16

fix: fallback to fp16 when bf16 requested but GPU doesn't support it (T4/Pascal)

fix: resolve CUDA OOM on T4 by using bfloat16, device_map=auto, and enabling gradient checkpointing

fix: revert device_map=auto (AMP gradient conflict on T4, use explicit cuda device instead)

fix: disable gradient checkpointing (conflicts with LoRA on T4, breaks backward pass)

fix: switch bf16->fp16 (Kaggle T4/P100 is Pascal, no Ampere bf16 support)

fix: add numpy to pip install deps for device_map=auto CPU offload

fix OOM: use device_map=auto with CPU offload + float16

fix: reduce memory for T4 GPU

fix: add missing datasets import in train_simple_nobnb.py

revert bitsandbytes: CUDA 13 runtime missing on Kaggle

fix OOM: enable 4-bit NF4 quantization + bf16 training

fix: force PyTorch 2.2.0+cu118 for P100 (sm_60) GPU compatibility

fix: use explicit cuda device map and use_reentrant=False for gradient checkpointing

fix OOM: batch_size 2->1, gradient_accumulation 4->8 for T4 GPU