hamlab-princeton/pastcasting-talkie-cpt

Pastcasting Talkie: Base, half-pass and full-pass checkpoints

One repository for the Talkie baseline and historical continued-pretraining checkpoints. Pass counts refer to the newly introduced period's corpus, not to lifetime training exposure. These are research checkpoints; availability does not imply that acquisition or retention acceptance gates passed.

Model	Folder	Previous label
Talkie 1930 Base	base/talkie-1930	Base
1931–1935 CPT, half-pass, step 3,372	cpt/1931-1935/half-pass-step-3372	“25%” of the former two-pass plan
1931–1935 CPT, one pass, step 6,736	cpt/1931-1935/full-pass-step-6736	“50%” of the former two-pass plan
1936–1941 CPT, half-pass, step 3,809	cpt/1936-1941/half-pass-step-3809	Half-pass
1936–1941 CPT, one pass, step 7,612	cpt/1936-1941/full-pass-step-7612	Full-pass

Upload verification status is recorded in release-status.json. The 1936–1941 one-pass checkpoint (step 7,612) has been uploaded and verified. No unfinished checkpoint is presented as available. Future period checkpoints use the same hierarchy.

The 1936–1941 half-pass precedes the expanded cumulative-replay continuation. That continuation resumed from step 3,809 and changed replay from step 3,810 onward; its final checkpoint is not identical to the original replay recipe continued unchanged.

Download one checkpoint

from pathlib import Path
from huggingface_hub import snapshot_download

subfolder = "cpt/1931-1935/half-pass-step-3372"
root = Path(snapshot_download(
    "hamlab-princeton/pastcasting-talkie-cpt",
    allow_patterns=[subfolder + "/*", "runtime/*", "base/talkie-1930/vocab.txt"],
))
checkpoint = root / subfolder

The upstream Base is preserved in its original final.ckpt format. CPT weights use FP32 sharded safetensors. These use the Talkie architecture and accompanying custom loader; the repository root is a collection, not a single AutoModelForCausalLM.from_pretrained target.

Install runtime/requirements.txt, then:

import sys
sys.path.insert(0, str(root / "runtime"))
from pastcasting.model import load_talkie_any, get_talkie_tokenizer
model = load_talkie_any(checkpoint)  # FP32; a 13B checkpoint needs substantial RAM
tokenizer = get_talkie_tokenizer(root / "base/talkie-1930/vocab.txt")

For Base, pass root / "base/talkie-1930/final.ckpt" to the loader. Available SFT adapters are stored under their matching checkpoint as adapter-1931-final, adapter-2027-final, and midpoint variants. They are adapters, not merged copies. The reported controlled factual comparisons use final seed 1931, with no seed selection. With the matching model already loaded, use peft.PeftModel.from_pretrained(model, checkpoint / "adapter-1931-final").

Factual dataset and interpretation

The complete updated 52,000-question factual bank contains 1,000 questions for every year 1921–1972, labels, source provenance, and independent reference calibration. It is distinct from the forecast-question generation pilot. Missing checkpoint evaluations are not zero scores.

The baseline comes from talkie-lm/talkie-1930-13b-base, pinned to b7c97680791f7fca4262c3c80b36ff7d666faab0. See NOTICE and LICENSE for upstream attribution. Raw model weights and SFT adapters are published here; optimizer/resume archives remain separate operational artifacts.