sy128/SpecComp-slot-sft-pretokenized-deepseekv4

SpecCompact slot-SFT pretokenized dataset (teacher: DeepSeek-V4)

Pretokenized SFT data for slot-wise agent-context compaction, built from scatyf3/speccompact-rollouts-deepseekv4 by src/train/scripts/build_llamafactory_slot_sft.py (speculative-compaction repo, llama-factory branch). Slot targets were generated by a DeepSeek-V4 teacher; sequences are tokenized with the Qwen/Qwen3-0.6B tokenizer (shared by all Qwen3 sizes), chat template applied with enable_thinking=False.

Format

datasets.DatasetDict saved with save_to_disk — not loadable via load_dataset; download and use load_from_disk, or point LLaMA-Factory's tokenized_path at the folder to skip its preprocessing entirely.

Columns: input_ids (int32, full chat-templated sequence), labels (int32, -100 over the prompt span, real ids over {slot_target}<|im_end|>\n), from (string source split: tau3 | appworld | liveclawbench).

Splits

95:5 holdout by md5(trace_id) hash — the same trace lands on the same side in every sibling dataset built this way (safe cross-teacher comparison).

split	rows	tau3	appworld	liveclawbench
train	54,454	15,569	516	38,369
validation	2,992	845	24	2,123

2 rows dropped (slot_target > 1024 tokens); no context truncation (cutoff_len=16384). See build_manifest.json for parameters and token-length percentiles, sample_previews.txt for decoded examples.

Usage (LLaMA-Factory)

model_name_or_path: Qwen/Qwen3-0.6B
stage: sft
tokenized_path: /path/to/this/folder