HZ0619/chrono-instruct-v1-20051231

ChronoGPT-Instruct (replication) — vintage 2005 (cutoff 2005-12-31)

Independent replication — not an official release. This model was fine-tuned by Huanyu Zhang as a from-scratch reconstruction of the training pipeline in He, Lv, Manela & Wu (2025). It is not produced or endorsed by the paper's authors (manelalab). The base weights and the SFT dataset are theirs; the instruction-tuning code and these fine-tuned weights are this replication's. Absolute numbers are not expected to match the paper to the decimal (see Limitations).

Model description

chrono-instruct-v1-20051231 is an instruction-tuned language model built by supervised fine-tuning (SFT) of the base model manelalab/chrono-gpt-v1-20051231 — a ~1.55B-parameter, 52-layer modded-nanoGPT U-net (26 encoder + 26 decoder layers with skip connections and value embeddings; model_dim 1536, 12 heads, vocab 50304, context 1792; RMSNorm, rotary position embeddings, QK-norm, ReLU² MLP, and logit softcap). The tokenizer is tiktoken GPT-2.

Chronological-consistency guarantee (the point). The base ChronoGPT vintage was pretrained only on timestamped text available before its knowledge cutoff τ = 2005-12-31, and this replication fine-tunes it on an instruction set that is screened to contain only pre-2000 instruction/response pairs (see Training data). Because pre-2000 ⊆ pre-τ for every vintage τ ≥ 1999, the fine-tuned model never saw any text created after 2005-12-31. This eliminates look-ahead bias / training leakage: the model is a point-in-time artifact that "knows" only what was knowable at its cutoff.

Why that matters. A model contaminated with future information silently inflates backtests of any text-conditioned prediction (return prediction, sentiment, event studies). A chronologically consistent model is safe to run as of its cutoff date, so a signal it produces on a 2005-12-31-or-earlier document could genuinely have been produced at that time.

Intended use

• Point-in-time (PIT) backtesting. Use the vintage whose cutoff precedes the timestamp of each document you score, so no evaluation ever uses a model that saw the future. The six released vintages (1999, 2005, 2010, 2015, 2020, 2024) let you assemble a leakage-free panel across time.
• Instruction following on pre-cutoff tasks (Alpaca-style single-turn instruction/response), and hidden-state embedding extraction for downstream text-as-data / asset-pricing research.
• Research into chronological consistency, temporal knowledge cutoffs, and instruction tuning of small time-stamped LLMs.

Out of scope

• Current-events or factual QA about anything after 2005-12-31. The cutoff blindness is intentional; the model cannot know post-cutoff facts and will answer as if they do not exist.
• Production chat / assistant deployment. This is a small (~1.55B) research model, not safety-tuned or RLHF-aligned; it can hallucinate and loop under greedy decoding.
• Multilingual, multi-turn, or long-context (>1792 token) use.
• Any high-stakes decision without independent validation.

How to load and generate

These repos store pytorch_model.bin + config.pt/config.json written by this replication's ChronoGPT.save_pretrained. Loading therefore uses the replication package's own ChronoGPT class (a lightly adapted, numerically bit-identical vendor of the authors' ChronoGPT_inference.py) — not transformers.AutoModel, and trust_remote_code is not required.

# Install the replication package (provides the ChronoGPT class + infer helpers)
pip install "git+https://github.com/zhanghuanyu0619/chrono-instruct-replication.git"
# or: git clone ... && cd chrono-instruct-replication && pip install -e .

import torch
from chrono_instruct import infer

# Loads weights from the Hub, moves to CUDA if available, sets eval mode.
model, device = infer.load("HZ0619/chrono-instruct-v1-20051231")

# The models were fine-tuned on the Stanford Alpaca prompt template
# (with a trailing newline after "### Response:"). Match it at inference time:
def alpaca_prompt(instruction: str, inp: str = "") -> str:
    if inp:
        return (
            "Below is an instruction that describes a task, paired with an input "
            "that provides further context. Write a response that appropriately "
            "completes the request.\n\n"
            f"### Instruction:\n{instruction}\n\n### Input:\n{inp}\n\n### Response:\n"
        )
    return (
        "Below is an instruction that describes a task. Write a response that "
        "appropriately completes the request.\n\n"
        f"### Instruction:\n{instruction}\n\n### Response:\n"
    )

prompt = alpaca_prompt("Name the President of the United States.")

# Greedy decoding by default (temperature=0.0, top_k=None) — matches the authors'
# ChronoGPT_instruct.py. A ~1.5B model can loop under pure greedy; for readable
# text turn on the anti-repetition guards (they change decoding, so leave them OFF
# to reproduce the paper's exact greedy output).
completion = infer.generate(
    model, device, prompt,
    max_new_tokens=128,
    temperature=0.0,          # 0.0 == greedy (argmax); set >0 to sample
    top_k=None,               # e.g. 50 to restrict sampling
    return_completion=True,   # return only the newly generated tokens
    # repetition_penalty=1.3, no_repeat_ngram_size=3,  # optional readability guards
)
print(completion)

For hidden-state embeddings (e.g. mean-pooled last-layer features for text-as-data):

vec = infer.embed(model, device, "Some pre-cutoff document text.", layer=-1, pool="mean")
print(vec.shape)  # (1536,)

Training data

• Dataset: manelalab/ChronoInstruct-SFT — Alpaca-format instruction/response triples from three sources (LLMs-from-scratch simple tasks → GPT-3 self-instruct → AllenAI Tulu-3 mixture), 647,944 rows as released.
• Temporal screen (the leakage guard). Only pairs the authors' GPT-4.1 temporal classifier labeled 0 ("knowledge available pre-2000") with confidence == 10 are kept — a deliberately strict double filter. This replication reproduces the paper's counts exactly: 647,944 → 425,119 retained pairs (1,097 / 67,136 / 356,886 across the three curriculum stages). A single pre-2000 screen is reused across all vintages because pre-2000 ⊆ pre-τ for every τ ≥ 1999.
• Loss masking. Response-only masked cross-entropy: prompt tokens are set to -100 so the loss scores only the response tokens.

Training procedure

• Full fine-tuning of all ~1.55B parameters with AdamW (no LoRA/PEFT).
• 3-stage curriculum, each stage resuming from the previous stage's weights:
  1. stage1_scratch (LLMs-from-scratch simple tasks) — 3 epochs
  2. stage2_self_instruct (GPT-3 self-instruct) — 2 epochs
  3. stage3_tulu (Tulu-3 mixture) — 2 epochs
• Hyperparameters: learning rate 3e-4 (all stages, per-stage cosine schedule, warmup 0.03, floor 0.1·lr); block size 1792; batch size 8 × grad-accum 4 (effective 32); gradient checkpointing on; global seed 123 (drives the train/val split, shuffle, and sampling).
• Compute: ~19.4 h wall-clock on a single 80 GB card (peak ~50.4 GB GPU) per vintage.
• Fidelity note: the paper releases the SFT data and weights but not the training code; the LR schedule and epoch counts here are this replication's tuning against the shape of the paper's Figure 1, not paper-exact values.

Evaluation

The intended evaluations follow the paper and are documented in this replication's report — see results/replication-report/README.md in the repository:

• SFT loss curves (Figures 1–2). ✅ Reproduced. Stage-3 validation cross-entropy falls monotonically as the cutoff advances (0.8691 at 1999 → 0.7855 at 2024), the paper's "later vintage = better language model, not leakage" structure.
• Chronological-consistency tests (Tables 2–3). U.S. president prediction and major-world-events tests: correct on pre-cutoff items, blind (0/N) on post-cutoff items. Coded and staged; run per vintage with chrono eval --repo HZ0619/chrono-instruct-v1-20051231 --cutoff 2005.
• AlpacaEval LC win-rate (Figure 3) vs Qwen-1.5-1.8B-Chat (paper reports ~54–62%). Coded; needs a judge API key.

Limitations & bias

• Small model (~1.55B). Limited reasoning/factual coverage; greedy decoding can loop (use the anti-repetition guards for readable output).
• Cutoff blindness is intentional, not a bug. The model has zero knowledge of anything after 2005-12-31 — that is the whole design. Do not use it as a general knowledge base.
• Not paper-exact. LR schedule and epoch counts are tuned to the shape of the paper's curves; absolute losses differ from the paper. The replication claim is qualitative structure (data-count match, clean curriculum chaining, monotone improvement with cutoff), not decimal parity.
• Packing. Examples are packed into 1792-token blocks; ~5.1% of Tulu examples are split across a block boundary (no response tokens dropped, but such examples are never seen whole in one forward).
• Inherits any biases of the base ChronoGPT pretraining corpus and the ChronoInstruct-SFT data. Not safety-tuned.

Citation

The paper this replicates:

@article{he2025instructchronogpt,
  title   = {Instruction Tuning Chronologically Consistent Language Models},
  author  = {He, Songrun and Lv, Linying and Manela, Asaf and Wu, Jimmy},
  year    = {2025},
  note    = {arXiv:2510.11677; SSRN 5348747}
}

Base models and SFT data (please cite the authors for both):

@article{he2025chronogpt,
  title   = {Chronologically Consistent Large Language Models},
  author  = {He, Songrun and Lv, Linying and Manela, Asaf and Wu, Jimmy},
  journal = {Working Paper},
  year    = {2025}
}

• Base model: manelalab/chrono-gpt-v1-20051231 (MIT license) — © manelalab.
• SFT dataset: manelalab/ChronoInstruct-SFT — © manelalab.

Model card authors

Replication and card by Huanyu Zhang. License MIT, inherited from the base model; the ChronoInstruct-SFT dataset is subject to its own terms — consult the dataset page before redistribution. This is an independent replication and is not an official manelalab release.