Ferriswheel1412/TemGenSAR-MovieLens25M

TemGenSAR — MovieLens-25M research checkpoints

This public research release contains the latest strict matched MovieLens-25M comparison from the 2026-09-08 server backup: Frozen SID, TASR-A, TASR-B, all three formal downstream seeds, and their shared t0 initialization. Prepared on 2026-10-15.

Source code: fferriswheel/TemGenSAR, implementation revision ce1ddd8511ef3bc863b90dcc3ac91c46ed1b3b68. Checkpoint files and their tokenizer/configuration files are unchanged from the backup. A derived catalog file contains only item mappings; user mappings are excluded.

Selection and results

TASR-B is the main research candidate, TASR-A is the frozen-codebook efficiency variant, and Frozen SID is the matched control. All formal seeds are retained rather than selecting the seed with the highest W2 test score.

Method	Rec H@10	Rec nDCG@10	Search H@10	Search nDCG@10
Frozen SID	0.452275	0.286285	0.706400	0.460273
TASR-A	0.452275	0.286483	0.706400	0.460248
TASR-B	0.452275	0.286628	0.706333	0.460220

These are means over downstream seeds 42 / 2024 / 2025, with W1 update/validation and W2 future testing. Evaluation uses 99 negatives and beam size 10. Rec evaluates 2,015 examples, Search 5,000. The catalog contains 18,191 items. Shared SID prefix and Search suffix stay fixed.

TASR-B improves mean Rec nDCG@10 by about 0.000344, or 0.12% relative, versus Frozen; all three seeds are positive. The saved analysis reports seed-level p=0.0303 and row-bootstrap p=0.0302. Rec H@10 is unchanged, and B versus A is not significant. This is evidence of a small ranking improvement, not a large overall retrieval gain. Search uses a fixed query/tag bank, so this benchmark does not establish natural Search drift.

TASR-A updates 14,528 SID parameters and changes 25 item codes. TASR-B updates 114,176 and changes 67. Both operational mappings have zero complete SID collisions. SID-stage timing in the original logs is about 15.75 / 21.88 seconds; these numbers exclude downstream T5 training.

The t0 model is the original shared initialization/teacher, trained with seed 2026. The formal temporal comparison excludes the separate development pilot seed 2026. The t0 model's older static validation scores are not directly comparable to W2 future-test metrics.

Layout

models/t0-p1-v13b/          Shared t0 T5-small initialization/teacher
models/frozen/seed-42/      Frozen SID, matched downstream update
models/frozen/seed-2024/
models/frozen/seed-2025/
models/tasr-a/seed-42/      TASR-A downstream T5 checkpoints
models/tasr-a/seed-2024/
models/tasr-a/seed-2025/
models/tasr-b/seed-42/      TASR-B downstream T5 checkpoints
models/tasr-b/seed-2024/
models/tasr-b/seed-2025/
sid/base/                  P1 anchor/suffix, lookup, operational mapping, GenSAR indices
sid/tasr-a/                Frozen-codebook adapter, operational mapping and indices
sid/tasr-b/                Aligned tokenizer checkpoint, operational mapping and indices
catalog/item_id_map.json   Raw MovieLens item ID <-> GenSAR integer index; no user mapping
evaluation/                Saved aggregate results and statistical comparisons
evidence/                  Phase 17–19 experiment notes
model_inventory.json       Roles, source paths, seeds, checkpoint metadata and provenance
validation.json            Local artifact validation and its scope
SHA256SUMS                 Checksums for release files except this checksum file itself

Each model folder contains the original model.safetensors, config.json, generation_config.json, tokenizer.json and tokenizer_config.json, plus a derived checkpoint-metadata record. The model has vocabulary size 33,639 and hidden size 512. Operational SID mappings use raw MovieLens IDs; the exported GenSAR index files instead use integer IDs. Use catalog/item_id_map.json when converting between them.

Loading

The server export records Transformers 5.9.0, PyTorch 2.6.0+cu124 and Python 3.12. Use a compatible isolated environment. Root-level auto-loading is intentionally not configured because this repository contains multiple checkpoints; specify a subfolder.

Public downloads do not require login. The pinned revision below contains the original model weights; documentation-only updates do not change the checkpoints:

from transformers import AutoModelForSeq2SeqLM, AutoTokenizer

repo_id = "Ferriswheel1412/TemGenSAR-MovieLens25M"
revision = "947652476014c1a04ba6b3d59e27896b7adc804d"
subfolder = "models/tasr-b/seed-42"  # fixed representative seed, not test-selected

tokenizer = AutoTokenizer.from_pretrained(
    repo_id, revision=revision, subfolder=subfolder, token=False, use_fast=True,
)
model = AutoModelForSeq2SeqLM.from_pretrained(
    repo_id, revision=revision, subfolder=subfolder, token=False,
)

For Transformers 4.52.3, the saved v5 extra_special_tokens list is incompatible with the older dict-shaped argument. The original files remain unchanged; a verified local load used these explicit tokenizer arguments:

tokenizer = AutoTokenizer.from_pretrained(
    local_checkpoint_directory,
    use_fast=True,
    local_files_only=True,
    extra_special_tokens={},
    additional_special_tokens=[f"<extra_id_{i}>" for i in range(100)],
)

This is a generative retrieval research checkpoint, not a general conversational model. Meaningful retrieval requires the original task prompts, history preprocessing, the correct method-specific SID mapping and constrained decoding/evaluation from the code repository. The model bundle does not contain raw interactions, training examples, user mappings, per-example predictions, training logs, or credentials.

Validation and limitations

validation.json records checks of checkpoint structure, all tokenizer vocabulary sizes, item/index mapping correspondence, operational SID uniqueness, preserved prefix/Search codes, and a representative local T5 load/forward pass. Historical ranking metrics were read from saved experiment outputs; the full benchmark and training were not rerun during packaging.

Underlying T5, GenSAR and MovieLens resources retain their respective licensing and data-use requirements. This upload does not introduce a new blanket license or authorize unrestricted redistribution of third-party data. The repository was made public on 2026-10-17; this visibility change does not alter model weights or expand third-party usage rights.

MovieLens data source and terms: GroupLens MovieLens-25M. Dataset citation: F. Maxwell Harper and Joseph A. Konstan (2015), The MovieLens Datasets: History and Context. Obtain original data from GroupLens; this release includes no raw ratings, tags or user histories.