Chronicle

A multimodal foundation model for joint language and time series understanding

---

Chronicle is a compact 324M-parameter decoder-only transformer trained from scratch on natural language and time series within a single unified architecture. Text tokens and time series patches share the same transformer blocks, attention mechanism, and residual stream — cross-modal capability emerges from shared parameters rather than from bolting a time-series encoder onto a pretrained LLM.

From one backbone, Chronicle:

• matches Gemma-3-270M-PT across 19 NLU benchmarks,
• sets a new bar for frozen-embedding time series classification on 24 UCR/UEA datasets,
• beats every supervised fusion baseline on Time-MMD multimodal forecasting.

📄 Paper: Chronicle: A Multimodal Foundation Model for Joint Language and Time Series Understanding (Quinlan, Levasseur, Li, Zhu)

🤗 Checkpoints: huggingface.co/InertialAI/Chronicle — both stages, the tokenizer, and the reference PyTorch implementation in one repo. Stage 1 (stage-1/) is the unimodal-batch pretrained backbone; Stage 2 (stage-2/) adds the short cross-modal alignment stage.

☁️ Hosted (coming soon): the same models will power InertialAI's API, including self-serve finetuning with automatic before/after evals and one-click GPU deployment.

Architecture at a glance

Parameters	324M, 16-layer decoder-only transformer
Width / heads	d=1024, 8 GQA heads (4 KV)
Modalities	text tokens + standardized time-series patches, one shared stream
Pretraining	unimodal batches (≈92% text / 8% TS) + short interleaved alignment stage
Objective	causal next-token / next-patch prediction

Install

git clone https://github.com/InertialAI/Chronicle.git
cd Chronicle
uv sync          # or: pip install -e .

Use the model locally

The checkpoints ship as safetensors with a minimal inference implementation (model.py + tokenizer.py, just torch + tiktoken) bundled in the HF repo. Verified example (text generation, CPU):

import json
import sys

import torch
from huggingface_hub import snapshot_download
from safetensors.torch import load_file

repo = snapshot_download("InertialAI/Chronicle")
sys.path.insert(0, repo)

from model import Chronicle, ChronicleConfig
from tokenizer import ChronicleTokenizer

config = ChronicleConfig(**json.load(open(f"{repo}/stage-2/config.json")))
model = Chronicle(config)
state = load_file(f"{repo}/stage-2/model.safetensors")
state = {k: v.float() if v.dtype is torch.bfloat16 else v for k, v in state.items()}
model.load_state_dict(state, strict=True)
model = model.float().eval()
model.cos, model.sin = model.cos.float(), model.sin.float()

tokenizer = ChronicleTokenizer.from_directory(f"{repo}/tokenizer")
ids = [tokenizer.get_bos_token_id()] + tokenizer.encode("Time series forecasting is")
with torch.no_grad():
    for _ in range(16):
        out = model(torch.tensor([ids]))
        logits = out[0] if isinstance(out, tuple) else out
        ids.append(int(logits[0, -1].argmax()))
print(tokenizer.decode(ids[1:]))
# -> "Time series forecasting is a technique used to forecast future values
#     of a time series based on historical data."

A transformers-native port (AutoModel.from_pretrained with forecast() / embed() conveniences, as used in examples/quickstart.py) is in progress; the examples target that interface, and the hosted API (coming soon) will serve the time-series pathway.

Finetune it locally

Each script is self-contained, downloads public data, and runs on a single GPU (or CPU for the small datasets):

Task	Script	Data
Forecasting	examples/finetune_forecasting.py	AirPassengers (public)
Classification	examples/finetune_classification.py	UCR GunPoint via aeon
Regression	examples/finetune_regression.py	synthetic sensor windows
Text generation (text and series+text)	examples/finetune_text_generation.py	AG News topics · EIA gasoline summaries
Text classification	examples/finetune_text_classification.py	SST-2 sentiment (GLUE)
Joint text + TS	examples/finetune_joint.py	captioned series

data/prepare_datasets.py builds ready-to-train JSONL datasets from public sources — ETTh1 change forecasting, EIA weekly gasoline-price change forecasting (via Time-MMD), series-grounded gasoline summary generation, and AG News topic naming. The records use the same shape as InertialAI's finetuning API, so the same files work locally and in the managed flow.

Two finetuning modes are shown, mirroring the paper's evaluation protocol:

• Linear probe (LP) — freeze the backbone, train a head on mean-pooled embeddings. This is the protocol behind the UCR/UEA and Time-MMD numbers.
• LoRA — low-rank adapters on the attention projections for the cases where a probe plateaus.

Prefer not to run GPUs yourself? The InertialAI API (coming soon) will offer the same finetuning as a managed flow — upload a JSONL, get a before/after eval, deploy.

Reproducing paper results

reproduction/ contains scripts and the expected numbers for a subset of the paper's benchmarks (UCR/UEA classification and Time-MMD forecasting), using only public data. From the paper (mean over 5 seeds, aeon default splits, linear probes on frozen embeddings):

UCR dataset	Chronicle Stage 1 (acc)	Chronicle Stage 2 (acc)	Best frozen TSFM baseline
GunPoint	0.919	0.851	0.931 (Moirai-2)
Coffee	0.893	0.893	0.964 (Moirai-2)
ECG200	0.846	0.848	0.840 (TimesFM)
FaceFour	0.864	0.759	0.609 (TimesFM)
Trace	0.936	0.848	0.802 (Moirai-2)

Time-MMD (MAE ↓, linear probe)	Chronicle Stage 2	Best fusion baseline
Energy	0.370	0.396 (GPT2+TimesFM)
Public Health	0.690	0.765 (BERT+TimesFM)
Social Good	0.399	0.441 (BERT+TimesFM)
Average NMAE (9 domains)	0.514	0.588 (GPT2+Moirai-2)

See reproduction/README.md for the full protocol and the per-dataset commands.

Citation

@article{quinlan2026chronicle,
  title={Chronicle: A Multimodal Foundation Model for Joint Language and Time Series Understanding},
  author={Quinlan, Paul and Levasseur, Jeremy and Li, Qingguo and Zhu, Xiaodan},
  journal={arXiv preprint arXiv:2605.20268},
  year={2026}
}

License

Apache 2.0 — see LICENSE.