Update README — v2 selector-aligned dataset, add pycryptodome, grounded training explanation

v2: Selector-aligned dataset — no guessing. Functions matched by 4-byte selector in bytecode dispatch table. Refusal examples for unmatched TAC. Grounded system prompt.

Update README — add dataset build step (Cell 0), update workflow for pre-built dataset

Update run_full.py — support pre-built dataset, faster startup

Add pre-built dataset support (USE_PREBUILT_DATASET flag) + keep on-the-fly fallback

Add standalone dataset builder script

8-bit training with BestLossUploader that cleans old checkpoints

8-bit training: load_in_8bit + liger_kernel + paged_adamw_8bit + batch=2"

Add PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True to fix fragmentation OOM

Fix OOM: batch=2 grad_accum=4, expandable_segments, r=32 needs more headroom"

Fix OOM: batch=4, grad_accum=2 (proven to work on 95GB with r=16, r=32 uses slightly more)"

Remove auto-upgrade (causes import conflicts). Install from source in Cell 2 instead."

Fix OOM: batch 8→6, grad_accum 1→2 (effective batch 12, fits ~90GB)

Fix: install transformers from source for qwen3_5 support, auto-upgrade in run_full.py

Fix Cell 1: total_mem → total_memory

Fix AttributeError: total_mem → total_memory

Full README with Colab notebook cells, VRAM breakdown, OOM recovery guide

Update train.py: AEON uncensored base, batch=8, r=32/α=64, ~88GB VRAM target

Update run_full.py: AEON uncensored base, batch=8, r=32/α=64, VRAM logging

Switch base model to AEON-7/Qwen3.6-27B-AEON-Ultimate-Uncensored (uncensored variant)

Add run_full.py: Colab wrapper with BestLossUploader for full 50K training

Update train.py: improved docstring, trackio removed (native integration via run_full.py), push_to_hub=False for Colab

Update: 1 epoch, batch_size=4, grad_accum=2 (2x faster, no overfit)

Fix: require transformers>=4.57.1 for qwen3_5 architecture support"

Add model card with full documentation

TAC to Huff + Solidity multi-target training script with full pipeline

Update: TAC → Huff + Solidity multi-target training script

Add training script for EVM bytecode to Solidity decompilation

Upload requirements.txt

initial commit