hamlab-princeton/wikispeedia-checkpoints-2026-10

Wikispeedia SFT and RL checkpoint archive

This public research archive preserves verified SFT and RL model exports from Wikispeedia experiments. Immutable, completed checkpoint exports from ongoing studies are added as they become available. Availability is recorded per checkpoint: catalog.json lists the main inventory; legacy-catalog.json lists the additional stopped run exports. Both catalogs mark verified availability. A planned entry is usable only when available is true.

Each checkpoint preserves its original Transformers or PEFT files. Catalog entries with storage_format: hf-byte-xor-zstd-v1 must first be restored using restore.py; other entries are directly loadable beneath checkpoints/<study>/<original-relative-checkpoint-path>. Compact archives use lossless byte XOR and Zstandard against retained full reference shards; restoration checks every original SHA-256. There is no quantization, merging of experimental arms, or conversion of model weights during archival.

Find and restore a checkpoint

1. Open catalog.json and select an entry whose available field is true.
2. Its subfolder locates the checkpoint and its revision pins the verified commit.
3. metadata/<id>.json lists every original file size and SHA-256 checksum.

To download and verify all files into an ordinary local model directory:

pip install huggingface_hub numpy zstandard
python restore.py --id CHECKPOINT_ID --destination ./checkpoint

For catalog entries without storage_format, Transformers accepts the full model directory directly. For compact entries, restore first and load the resulting local directory:

from transformers import AutoModelForCausalLM, AutoTokenizer
repo = "hamlab-princeton/wikispeedia-checkpoints-2026-10"
model = AutoModelForCausalLM.from_pretrained(repo, subfolder=entry["subfolder"], revision=entry["revision"])
tokenizer = AutoTokenizer.from_pretrained(repo, subfolder=entry["subfolder"], revision=entry["revision"])

Historical PEFT adapters retain their original configuration, including a local base-model path. Load Qwen/Qwen2.5-7B-Instruct explicitly as the base model, then load the restored adapter with PEFT. Original checkpoint files are preserved exactly.

Experimental families

• no-interface-training-prompts-20261013/do-not-edit: do-not-edit training, with editing tools still available. Completed SFT exports are uploaded during the ongoing four-epoch stage; selected SFT initializers and downstream RL exports are added as they become available. This is distinct from do-not-edit evaluation of neutral-trained models.

• no-interface-reasoning-20261009: original-data SFT, CoT on/off, goal/shortest-path success filtering; also an earlier RL protocol, kept separate.

• no-interface-navigation-only-20261010: matched navigation-only SFT arms without editing demonstrations.

• no-interface-rl-fp32-20261010: corrected RL protocol, with arrival and shortest-path rewards; the folder structure identifies the SFT initializer and RL reward.

• Other study directories preserve older browser experiments, continuous SFT, pilots, and the explicitly taught navigation warm start. These are different protocols and should not be pooled with the no-interface study.

The main research question is whether RL increases editing/shortcut use or improves navigation without editing, and how neutral versus do-not-edit training changes that behavior. Model availability is not evidence of improved performance. See the dated experiment report for evaluations and protocol qualifications; that report may require owner access.

This archive contains model exports, not training datasets, raw model conversations, credentials, or optimizer/RNG resume state. Selected/final local checkpoints still required by live training and retention jobs remain cached locally. Other local model copies are released only after remote file identities match the originals. Original studies and checkpoint step names are preserved for provenance.

Additional legacy run exports

legacy-catalog.json records 14 older checkpoint folders discovered under runs/, including 3B/7B smoke runs. Three entries reuse the identical already-verified checkpoint in the main catalog. The other eleven have their own checkpoints/runs/... folders. These are historical diagnostics, not additional arms of the current 7B study.

python restore.py --catalog legacy-catalog.json --id CHECKPOINT_ID --destination ./legacy-checkpoint

Built with Qwen. All 7B-derived checkpoints use Apache-2.0. Only the explicitly named qwen25-3b historical smoke checkpoints use the Qwen Research License. The complete applicable licenses and attribution notices are included in LICENSE and NOTICE, with separate copies inside the additional run folders. The weight files are fine-tuned derivatives; archival preserves their original checkpoint bytes.

Evaluation data and saved reasoning

The companion Wikispeedia evidence dataset preserves SFT examples, checkpoint evaluations, RL training rollouts and saved model reasoning. Its checkpoint-index.json links model checkpoints to evidence; rl-lineage-index.json separates reward branches and their SFT initializers. Original trainer token IDs and readable decoded responses are both retained. Evidence uploads are ongoing: consult the dataset catalog for verified availability.

Organization ownership

Transferred to hamlab-princeton on October 20, 2026. All existing checkpoint files, commit revisions, and verification hashes were preserved. The previous od2961 repository address redirects here.

Lossless storage

138 checkpoints use verified lossless archives. Every original file is recovered byte for byte; weights are not quantized or converted. Each uploaded compressed shard was downloaded and fully decoded before its redundant full copy was retired. Full reference shards are retained under archive/references/. Original metadata and checksums remain under metadata/.

Use the current restore.py and catalog. Historical raw download revisions were retired with the owner's approval; archive/provenance/revision-map.json maps them to current restore entries. archive/provenance/object-map.json maps every original large-file checksum to retained bytes or a verified reconstruction recipe. archive/provenance/original-history.bundle preserves the original Git metadata and small-file history; large files are recovered using the object map.