Prepared response-targeted SFT and causal replay for LiLM1-Tool-234M.
Recipe: lilm1-posttraining-sft-200m-v1
Train sequence tokens: 200,007,204
Train assistant/causal target tokens: 51,891,502
Tokenizer: HuggingFaceTB/SmolLM2-135M@93efa2f097d58c2a74874c7e644dbc9b0cee75a2
Maximum length: 4,096
Read dataset_manifest.json before use. Tool examples come from the verified
structured record_json field, not the original plain-completion rendering.
The… See the full description on the dataset page:
https://huggingface.co/datasets/glouriousgautam/lilm1-posttraining-sft-200m-v1.
