processed-new-cl-fewrel

FewRel continual relation extraction at the sentence level, 5 permutations x 10 tasks, already tokenised for Qwen3-0.6B, so nothing has to be built or processed after downloading. One archive, fewrel_all.tar.gz; unpack it at the root of the OpenED repo (scripts/qwen/cre_sent/fetch_data.sh fewrel does both).

This replaces the pair-level splits of datht/processed-cl-fewrel. There, a row was one entity pair and the prompt named the pair. Here an input is a sentence, the prompt names no entities, and the model writes every triple it finds.

• Built by tools/build_cre_sent.py from the full WAVE pool (80 relations x 700 instances). Instances that share a sentence are merged and duplicate triples dropped, so the triples of a sentence are its full annotation.
• 3 sentences are dropped because their full row does not fit the length limits of the pipeline: 460 prompt tokens, 768 in total, and 308 generated tokens (generation left-pads the prompt to 460, and a cut-off JSON scores the whole row as empty). That leaves 55,952 triples in 54,136 sentences, 1,679 of them with more than one triple.
• Train/dev/test are split by sentence, so no sentence sits in two splits. The RP-CRE sizes are kept as caps per relation, 420/140/140. Whole sentences are placed, so train holds 393 to 420 triples per relation; dev and test hold exactly 140.
• Each permutation draws its own partition and order of the relations (seed 2021 + 100p), 8 relations per task, as in RP-CRE.
• Task t, train.jsonl: the training sentences with a triple of the current task. Their target keeps only those triples, so triples of earlier tasks are stripped (task-wise annotation). It also holds the memory: 10 sentences per earlier relation, whose target has every triple with a label seen so far. Over tasks 1-9, 1.7% of the new training sentences hold a stripped triple.
• Train rows carry "is_memory": true|false. A memory target can also hold a triple of the current task, so the labels alone cannot tell memory rows apart. tools/process_data.py copies the field into the processed jsonl, and the trainer, replay oversampling and pseudo-labelling read it.
• Task t, dev.jsonl / test.jsonl: cumulative. Every sentence with a triple of a label seen so far, with all of those triples.
• One record per triple, ordered by subject and then object position, in the CED schema so the same evaluator runs unchanged: [<subject span>, <relation type>, [[<object span>, "object"]], <description>]. Full triple F1 is the argument score of ed_eval.py. Its trigger score only checks (subject, relation).

Layout mirrors the OpenED repo, so each file goes to the same relative path:

• data/fewrel_sent_perm{0..4}/streams.json: the relation types of each task, in stream order
• data/fewrel_sent_perm{0..4}/{0..9}/{train,dev,test}.jsonl: the task split ({system_prompt, user_prompt, response}, plus is_memory on train rows). The pseudo-labelling and self-distillation stages read these, so they are needed even though the tokenised copies exist.
• processed_data/fewrel_sent_perm{0..4}/{0..9}/qwen/: tools/process_data.py output (train_0.bin / .idx, plus valid and test), built with --max-prompt-length 460 --t-max-prompt-length 640 --dev-num 1000.

The directories end in _sent_perm, so unpacking does not overwrite the old fewrel_perm* splits. Train with scripts/qwen/cre_sent/ (run names carry _sent); the old scripts/qwen/cre/ runners still read the pair-level data.