pipeline-error-propagation-dialogue-summarization-led-baseline-seed73

LED-BASE — baseline, training seed 73

Selected checkpoint from the revised experiment accompanying Controlled Error Propagation Analysis in Transformer-Based Pipelines for Dialogue Summarization. Release: 2026-10-20. Master replication record (all versions): 10.5281/zenodo.21792310.

Full-meeting baseline, with the complete dialogue tokenized and truncated to the input limit.

Training and checkpoint selection

The experiment uses 94 training, 20 validation and 19 test AMI scenario meetings. training_logs/ contains the configuration, complete recorded history, probe indices, token statistics, completion marker and selected-checkpoint summary. The supplied files preserve the final trained weights and tokenizer. Optimizer and RNG state for resuming training are not part of this selected checkpoint.

Loading

from transformers import AutoTokenizer, AutoModelForSeq2SeqLM

repo_id = "Immich/pipeline-error-propagation-dialogue-summarization-led-baseline-seed73"
tokenizer = AutoTokenizer.from_pretrained(repo_id)
model = AutoModelForSeq2SeqLM.from_pretrained(repo_id)

Use Transformers 4.57.6, as recorded by the experiment. For exact reproduction, pin the Hub commit recorded in the new replication package's model registry. Use the publication notebook's complete generation routine, including input construction, oracle/noise manifests, allocation, concatenation and meeting-level scoring. Loading the weights alone does not reproduce the experimental pipeline.

Scope and attribution

These models were evaluated on English AMI scenario meetings about remote-control design. ROUGE measures lexical overlap and does not establish factual accuracy. Oracle-minus-baseline differences include privileged reference-derived inputs; they do not isolate an architectural advantage of the pipeline. A shared token cap does not imply equal realized summary lengths across models or approaches.

Repository contributions retain the MIT metadata of the previous model release. Pretrained components retain their upstream Apache-2.0 terms and attribution; the original licence and model card are included as UPSTREAM_LICENSE.txt and UPSTREAM_MODEL_CARD.md. AMI annotations retain their CC BY 4.0 attribution: AMI Meeting Corpus, AMI Consortium.

Weight SHA-256: 4319da984137d7e41c8f898160aa364a59265169d509727068437915753dd44c. See RELEASE.json for the exact backbone revision and release identifiers.