t4llm β source backup
This repository is a source-only backup of the crash-only, resumable 0.6Bβ1.7B LLM training pipeline for a single free T4. It contains no model weights, no training data and no checkpoints. No licence has been chosen for the code or the eventual model weights; do not redistribute the merged model until one is.
- Canonical source: https://github.com/Ggboykxz/t4llm
- Current backup commit, archive and bundle names, and SHA-256 digests:
backup/LATEST.json - Archives and full git bundles:
backup/
Restore
git clone https://github.com/Ggboykxz/t4llm.git
# or, from the newest bundle listed in backup/LATEST.json:
git clone backup/<bundle-name>.bundle t4llm
The shipped pipeline is fp16-only (SDPA, no FlashAttention, no bf16) and
supports CPT, prompt-masked SFT and sigmoid DPO with crash-only checkpointing.
Measured smoke tests include a merged fp16 safetensors + f16 GGUF export on
Qwen3-0.6B-Base; see docs/BENCH.md after restoring. One step is not a trained
model and no benchmark is claimed for a real domain run.
Inference Providers NEW
This model isn't deployed by any Inference Provider. π Ask for provider support