BlueSourceJY/ERCBench-StoryMem-400

Independent server batches

The complete 1–400 ranking is unchanged. Assign the existing server 0–100 and each new server one additional batch. Ranges use zero-based, end-exclusive offsets: 100–150 contains 50 episodes, ranks 101–150. The seven batches have no overlaps and cover all 400 episodes. A batch priority expresses the preferred expansion order; separately assigned servers may run their batches concurrently. No central scheduler is included.

Priority	Offset range	Existing ranks	Episodes	Shots	Archive
P1	0–100	1–100	100	500	ZIP
P2	100–150	101–150	50	252	ZIP
P3	150–200	151–200	50	248	ZIP
P4	200–250	201–250	50	252	ZIP
P5	250–300	251–300	50	248	ZIP
P6	300–350	301–350	50	252	ZIP
P7	350–400	351–400	50	248	ZIP

Download only the assigned batch (example: 100–150)

hf auth login
hf download BlueSourceJY/eec-bench-storymem-400 batches/storymem_inputs_0100_0150.zip batches/storymem_inputs_0100_0150.zip.sha256 --repo-type dataset --local-dir storymem_dataset
cd storymem_dataset/batches
sha256sum -c storymem_inputs_0100_0150.zip.sha256
unzip storymem_inputs_0100_0150.zip
python storymem_inputs_0100_0150/verify_bundle.py
python storymem_inputs_0100_0150/list_story_scripts.py > assigned_story_scripts.txt

Replace 0100_0150 with the assigned batch ID. Each ZIP is self-contained: its index.json and generation_queue.jsonl contain ONLY that batch, with portable paths resolved relative to its extracted directory. Feed each listed native story_script.json to the existing StoryMem runner through its official --story_script_path argument. Use a separate output directory per episode. list_story_scripts.py only enumerates inputs.

Full bundle and compatibility

Download only storymem_inputs_400.zip and its .sha256 sidecar if you need all 400. The full archive extracts to storymem_inputs/. Select one batch using python storymem_inputs/list_story_scripts.py --batch 0100_0150 or use the corresponding batches/batch_0100_0150.jsonl. Existing core_100.jsonl, priority_200.jsonl, stage2_additional_100.jsonl, stage3_remaining_200.jsonl, index.json and episode files remain byte-identical. The old coarse stages remain compatibility views. New priority_150/250/300/350.jsonl are CUMULATIVE evaluation subsets, not additional jobs. Do not dispatch a cumulative subset as a new batch. All models share the same episode IDs/order. These archives contain StoryMem inputs; HoloCine requires its own native input adapter and runtime. The manifest/ID lists define common assignments for both.

Preserved input semantics

Every episode includes unchanged story_script.json, contract.json, alignment_contract.json, entity_schedule.json and shot_mapping.json. Use the native video_prompts as delivered, with all cut flags true. Earlier context is already included where needed; do not append schedules, claims or other shots to the text. Map native scene/shot output names using shot_mapping.json; scene occurrence numbers are not persistent scene identities. Schedule omission is not a prohibition. Original ambiguity records and the previous validation report are retained in the full archive. Per-batch ZIPs retain the original contracts and claims verbatim.

This is a dispatch/export change only: no reranking, new prompt rewriting, model inference or fresh semantic annotation. Full source order, first100 and all episode payload hashes are verified. The previously audited tokenizer inputs are unchanged; this update does not repeat tokenization. Lower batch number means earlier priority, not a calibrated difficulty or quality score. Existing strict cluster campaign barriers were not changed by exporting these independent assignments.

For reproducible downloads, add --revision to hf download and use the same commit on all rented servers. Keep the account's existing private-repository access. Official StoryMem: https://github.com/Kevin-thu/StoryMem Previously tested code commit: 052c68d2627a22a95f79dbb4d3376cc30c96f1a3.