geodesic-research/pa-warm-start-sft-25b-rendered-review
pa-warm-start-sft-25b rendered review sample (n=200) 200 uniformly-sampled conversations from geodesic-research/pa-warm-start-sft-heavy-25b-mix (default/train, the control-pretraining 30B baseline SFT corpus), rendered EXACTLY as the training pack renders them: the library's _chat_preprocess (tool-call normalization + think-HISTORY chat template + assistant-only loss mask). Columns: rendered_text (the full string the model sees), trainable_spans_only (concatenation of… See the full description on the dataset page: https://huggingface.co/datasets/geodesic-research/pa-warm-start-sft-25b-rendered-review.
pa-warm-start-sft-25b rendered review sample (n=200)
200 uniformly-sampled conversations from geodesic-research/pa-warm-start-sft-heavy-25b-mix (default/train, the control-pretraining 30B baseline SFT corpus), rendered EXACTLY as the training pack renders them: the library's _chat_preprocess (tool-call normalization + think-HISTORY chat template + assistant-only loss mask).
Columns: rendered_text (the full string the model sees), trainable_spans_only (concatenation of loss-mask=1 spans — what the model is actually trained to produce), per-row stats (turns, reasoning coverage, trainable fraction, think-tag balance).
Sample trainable-token fraction 0.8183 vs 0.8182 measured on the full 25.0B-token pack — the sample and corpus agree to 4 decimal places. Built 2026-08-27 for manual data review.