-
Upload configs
-
Fix OOM: disable torch.compile, reduce batch sizes, double grad_accumulation
Faaz
-
Fix hidden_size: 4096 -> 3584 to match Qwen2.5-Coder-7B-Instruct
Faaz
-
Add WebSight vision data pipeline: download script, image-aware data loader, phase data routing
Faaz
-
Day 3 COMPLETE: Full model architecture
Faaz
Files added:
- src/model/architecture.py (Qwen2.5-Coder-7B + LoRA)
- src/model/vision_encoder.py (CLIP ViT-L/14, 256 tokens)
- src/model/fusion_layer.py (Linear+LayerNorm prepend)
- src/model/mindi_model.py (MINDI15 complete model)
- src/training/mindi_trainer.py (3-phase MI300X trainer)
- scripts/train.py (master training script)
- configs/training_config.yaml (MI300X config)
- setup_mi300x.sh (MI300X setup script)
- scripts/quality_filter.py, split_data.py, data_stats.py
- scripts/upload_everything_to_hf.py
Model: Qwen2.5-Coder-7B + LoRA + CLIP Vision
Ready for MI300X training!
-
Day 1 Complete: Tokenizer setup — Qwen2.5-Coder-7B base + 22 MINDI special tokens (vocab 151,685), wrapper class, full format test
Faaz
-
feat: initial project scaffold for MINDI 1.5 Vision-Coder
Faaz
- 21 directory structure (src, api, configs, scripts, data, etc.)
- 4 YAML configs (model, training, data, search)
- Core modules: model, vision encoder, agents, sandbox, search
- FastAPI server with auth middleware
- EnvLoader + ConfigLoader utilities with typed dataclasses
- Training pipeline, dataset loader, inference pipeline, evaluator
- Comprehensive validation script (validate_setup.py)
- .gitignore properly excludes .env, venv/, checkpoints/, etc.
