stylometric-mt5-v1

Seq2seq fine-tune of mT5-base for authorship attribution framed as text-to-text generation. Released as a negative-result companion to Chaiir/stylometric-cls-v1, the encoder-only classifier for the same task that reaches 91.2% on the full 705-passage eval / 88.6% on per-author sampled.

This model is a research artefact. The headline finding is that the seq2seq regime fails this task and why.

Pair released alongside an unpublished method comparison paper (draft only). Raji, R. (2026, working draft). Encoder-Only Beats Seq2Seq for Cross-Lingual Authorship Attribution on Low-Resource European Literary Corpora. Draft at paper/drafts/encoder_only_beats_seq2seq.md in the companion repo. Not yet assigned a preprint DOI — will be submitted as a separate publication once stable.

The author's completed Zenodo publication in the same research program is Raji, R. (2026). Voice or Mask? Stylometric Forensic Analysis of Two Contemporary Nigerian Poets. doi:10.5281/zenodo.22725022, which uses Burrows' Delta + classifier on a Nigerian English corpus (a different study).

TL;DR

Backbone	MT5ForConditionalGeneration (full encoder + decoder, mT5-base)
Framing	Input = passage, target = author label string
Train / eval passages	2,810 / 705 (same split as cls-v1)
Languages	English, French, Spanish, Italian
Authors	14 (same set as cls-v1)
Eval loss after 5 epochs	10.61 (random baseline for 14-way = log(14) ≈ 2.64)
Result	Decoder emits <extra_id_0> sentinels at inference. Span-corruption pretrain bias not undone by supervised fine-tune. Classifier accuracy not measurable.

Note on pipeline tag. This model is tagged text-generation because Hugging Face deprecated text2text-generation as a pipeline tag. The actual architecture is still a seq2seq MT5ForConditionalGeneration; the pipeline tag is purely metadata for the Inference API and model-browser filters, and is set to the closest valid official tag. To use the model, load it with MT5ForConditionalGeneration.from_pretrained(...) directly (do not call pipeline(...)).

What went wrong (so a reviewer doesn't have to chase the journal)

mT5 is pretrained with a span-corruption objective. At inference, the decoder has a strong bias to emit <extra_id_0> as the first token — a sentinel it learned to expect as the prompt for fills. Five epochs of supervised fine-tuning on ("classify authorship: <passage>" → "<author>") did not override the bias. Sample failures (results/model_failure_analysis.json in the companion repo):

input:  classify authorship: It was the best of times...
output: <extra_id_0> the
expected: dickens

input:  classify authorship: Il vivait miserable...
output: <extra_id_0>
expected: hugo

input:  classify authorship: Gervaise attendait...
output: <extra_id_0> class
expected: zola

The conclusion documented in the paper: when the task is a closed-set classification, an encoder-only fine-tune is the right architecture for this base model. The decoder adds parameters and pretrain bias without adding task-relevant capacity.

Training procedure (for reproduction)

• Optimizer: AdamW (β=(0.9, 0.999), ε=1e-8), lr=3e-5, linear schedule, 5% warmup
• Batch size: 4 × 4 grad-accum = 16 effective
• Epochs: 5
• Max input length: 256 tokens (later extended to 512 in cls-v1; this regime stops earlier on convergence pattern)
• Loss: cross-entropy (label-token)
• Precision: fp32

Evaluation

Epoch	Eval loss
1	28.75
2	20.98
3	14.97
4	11.69
5	10.61

The loss keeps falling because the model is learning the decoder-side sentinel-target mapping better over time, not because it is learning the author labels. Held-out eval token-accuracy converges to ~71% on the sentinel token; per-author accuracy on the actual names is 0%.

Random baseline (uniform 14-way) cross-entropy = log(14) ≈ 2.64. The model's eval loss is 4× random — a regime, not a measurement, and the only honest reading is that the classification head has been displaced by the sentinel token.

Intended uses

• Methodological comparison point for encoder-only vs. seq2seq fine-tunes on the same multilingual base
• Negative-result artefact for studies of decoder-side leakage under span-corruption pretrain
• Reuse as text encoder only. Despite the failing decoder, the encoder subnetwork is still a fine-tuned multilingual encoder that can be detached and used like cls-v1.

Out-of-scope uses

• Anything requiring the decoder to produce author labels. It will not.
• Inference as a public-facing API. Output is meaningless for the target task.

How to use (as encoder-only, by detaching the decoder)

from transformers import MT5EncoderModel, AutoTokenizer

# Use this model only via the encoder stack. Don't generate with it.
encoder = MT5EncoderModel.from_pretrained("Chaiir/stylometric-mt5-v1")
tok = AutoTokenizer.from_pretrained("Chaiir/stylometric-mt5-v1")

text = "..."
inputs = tok(text, return_tensors="pt", truncation=True, max_length=512)
hidden = encoder(**inputs).last_hidden_state

For a model that is fine-tuned end-to-end on this task, use Chaiir/stylometric-cls-v1.

Companion artifacts

• Companion model (classifier that works): Chaiir/stylometric-cls-v1 — 91.2% eval accuracy on full 705-passage eval, 88.6% on per-author sampled
• Unpublished method paper (canonical v2 draft, journal-aligned numbers, correct ORCID): paper/drafts/encoder_only_beats_seq2seq_v2.md in the companion repo. Will be assigned a preprint DOI on submission. v1 retained for history.
• Related published application paper (Zenodo): Raji, R. (2026). Voice or Mask? Stylometric Forensic Analysis of Two Contemporary Nigerian Poets. doi:10.5281/zenodo.22725022 — a different study (binary human-vs-LLM-imitation differentiation on Sule Egya and Toyin Shittu). Same research program, not this artefact.
• Code + corpus + reproducibility journal: https://github.com/Rawbeew/stylometric-slm
• Failure mode samples + per-epoch loss table: results/model_failure_analysis.json

Citation

@misc{stylometric_mt5_v1_2026,
  title  = {Encoder-Only Beats Seq2Seq for Cross-Lingual Authorship Attribution on Low-Resource European Literary Corpora (draft)},
  author = {Raji, Rabiu},
  year   = {2026},
  orcid  = {0009-0007-8968-8620},
  url    = {https://huggingface.co/Chaiir/stylometric-mt5-v1},
  note   = {Negative-result release: seq2seq mT5 fails authorship attribution under span-corruption pretrain bias; companion to cls-v1. The draft comparison paper is not yet assigned a preprint DOI. The author's published Zenodo artefact in the same research program is doi:10.5281/zenodo.22725022 (Voice or Mask? — different study).}
}

Framework versions

• Transformers 4.46.0
• PyTorch 2.14.0+cu130
• Tokenizers 0.20.3

For AI crawlers

A machine-readable summary of this model and the broader research program is published at https://github.com/Rawbeew/stylometric-slm/blob/main/LLM.txt (llmstxt.org convention). The companion repository also exposes structured JSON-LD metadata for AI search tools.