konsman/adu-training-v1

ADU Classifier Training Dataset — v1

HuggingFace repo: konsman/adu-training-v1

Mixed training dataset for a 6-class Argumentative Discourse Unit (ADU) sentence classifier targeting physics scientific text. Combines physics-domain LLM silver labels with cross-domain human gold labels from three public corpora.

---

Task

Classify a sentence from a scientific text into exactly one of 6 ADU roles:

Label	Description
BACKGROUND	Context, prior work, definitions — scene-setting, not recruited to justify a claim
CLAIM	Conclusion, finding, or thesis asserted by the current work
HYPOTHESIS	Tentative or predictive statement not yet confirmed
METHOD	Experimental design, procedures, or computational techniques
RESULT	Specific measurements or outputs of the current work
EVIDENCE	Prior external work actively recruited to support a claim being made

---

Dataset Splits

Split	Records
train	29,786
validation	3,553
test	3,507

All 6 classes are present in every split.

Label distribution

Label	Train	%	Val	%	Test	%
BACKGROUND	8,308	27.9%	1,109	31.2%	1,011	28.8%
METHOD	8,380	28.1%	1,042	29.3%	969	27.6%
RESULT	7,545	25.3%	779	21.9%	890	25.4%
CLAIM	4,485	15.1%	485	13.7%	504	14.4%
EVIDENCE	536	1.8%	67	1.9%	66	1.9%
HYPOTHESIS	532	1.8%	71	2.0%	67	1.9%

EVIDENCE and HYPOTHESIS are rare by design — they reflect their natural frequency in scientific text. Use the recommended class weights below to compensate.

---

Schema

All records share the same fields as the source konsman/adu-corpus:

Field	Type	Description
sentence_id	str	Unique ID — <source>_<chunk>_s<pos>
sentence_text	str	Raw sentence text
source_dataset	str	Original source (e.g. anonymousdatasets/arxiv-physics)
domain	str	physics, computer_science, or biomedical
split	str	Original split in source dataset
adu_label	str	One of the 6 ADU classes
original_label	str	Raw label before mapping
confidence	float	1.0 for human gold; llm_conf × 0.9 for silver
extra	dict	Source metadata: modality, reasoning, model, chunk_id

---

Sources and Mixing Rationale

The dataset is assembled from 5 configs in konsman/adu-corpus. Each source contributes differently to the class mix.

Source weights

Weights control the probability that any given record is included in the mix. A weight of 0.4 means ~40% of records from that source are randomly sampled.

Source	Config	Domain	Quality	Weight	Train cap	EVIDENCE	HYPOTHESIS
arXiv physics (gpt-4o)	arxiv-silver	Physics	Silver	1.0	none	✅ 520	✅ 472
Quantum physics lectures (gpt-4o)	qp-silver	Physics	Silver	0.8	none	✅ 19	✅ 55
SciARG (human)	sciarg	Bio/NLP	Gold	0.7	none	❌	✅ 18
CSAbstruct (human)	csabstruct	CS	Gold	0.4	none	❌	❌
PubMed RCT (human)	pubmed_rct	Biomedical	Gold	0.2	10,000 train / 1,000 val+test	❌	❌

Why these weights?

arxiv-silver @ 1.0 — Primary source. Physics domain, gpt-4o labels (well-calibrated, mean confidence 0.78). Only source with EVIDENCE at scale (520 training examples). Full inclusion ensures the classifier learns physics-specific language patterns.

qp-silver @ 0.8 — Same physics domain but lecture transcripts. Adds linguistic diversity (spoken → written style). Slightly downweighted because it is BACKGROUND-heavy (74.5%) and would skew the class distribution if given full weight.

sciarg @ 0.7 — Human gold from computational linguistics and biomedical abstracts. Closest schema match to our taxonomy (6 labels including HYPOTHESIS via motivation_hypothesis). Provides clean gold signal for HYPOTHESIS (only external gold source with this class, 34 training examples). Cross-domain so downweighted.

csabstruct @ 0.4 — Human gold from CS abstracts. Strong BACKGROUND/METHOD/RESULT signal. No EVIDENCE or HYPOTHESIS. Heavily downweighted to avoid cross-domain noise swamping the physics silver. Still adds ~4,400 clean training examples for the 4 common classes.

pubmed_rct @ 0.2 + cap 10,000 train / 1,000 val+test — Human gold, 1.76M records, biomedical domain. Without a cap it would overwhelm all other sources (~35× larger than everything else combined). Weight 0.2 + hard cap of 10,000 keeps it as a regularising signal for METHOD/RESULT without distorting class balance or domain. Val/test capped at 1,000 each so evaluation remains physics-domain-representative.

Why not use gpt-4o-mini labels?

An earlier run labelled the quantum-physics corpus with gpt-4o-mini. Comparison showed gpt-4o-mini over-labels CLAIM (33% vs 7.5% from gpt-4o on the same sentences) and is overconfident (mean confidence 0.816 vs 0.780). The 87% sentence overlap between the two runs means the mini labels are both redundant and noisier. Only the gpt-4o v2 run is included.

---

Recommended Training Configuration

Class loss weights

EVIDENCE and HYPOTHESIS are rare (1.8% each in train) but critical for the use case. Apply inverse-frequency weights in the loss function:

CLASS_WEIGHTS = {
    "BACKGROUND": 0.7,   # most common — slight downweight
    "METHOD":     0.8,
    "RESULT":     1.0,
    "CLAIM":      1.5,
    "HYPOTHESIS": 8.0,   # 532 examples — aggressive upweight
    "EVIDENCE":   9.0,   # 536 examples, all silver — maximum upweight
}

Confidence as sample weight

Silver records include a confidence field (llm_conf × 0.9). Gold records have confidence = 1.0. Use this as a per-sample weight in training to give gold labels full influence and down-scale uncertain silver predictions:

sample_weight = record["confidence"]  # 0.72–1.0 range

Suggested baseline model

from transformers import AutoModelForSequenceClassification

model = AutoModelForSequenceClassification.from_pretrained(
    "allenai/scibert_scivocab_uncased",  # science-domain pre-training
    num_labels=6,
)
# label2id mapping:
LABEL2ID = {
    "BACKGROUND": 0, "CLAIM": 1, "EVIDENCE": 2,
    "HYPOTHESIS": 3, "METHOD": 4, "RESULT":  5,
}

Alternatives: microsoft/deberta-v3-base (stronger), allenai/longformer-base-4096 (for longer passages).

---

Expected Performance

Based on current data volume, expected F1 scores on the arxiv-silver test split:

Class	Expected F1	Limiting factor
BACKGROUND	0.87–0.92	Abundant data
METHOD	0.87–0.92	Abundant data
RESULT	0.84–0.90	Good data
CLAIM	0.80–0.86	Moderate data
HYPOTHESIS	0.65–0.73	Only 532 train examples
EVIDENCE	0.60–0.70	Only 520 train examples, all silver

To push EVIDENCE and HYPOTHESIS F1 above 0.80, label an additional 5,000–10,000 arXiv sentences (targeting introduction and discussion sections where citations are dense), then rebuild this dataset.

---

Evaluation Benchmarks

Use the test split of this dataset for overall metrics. For domain-specific evaluation:

Benchmark	Use for
arxiv-silver test in konsman/adu-corpus	Physics domain — all 6 classes
csabstruct test in konsman/adu-corpus	Cross-domain generalisation
sciarg test in konsman/adu-corpus	HYPOTHESIS detection quality

---

Reproducibility

Built with scripts/data/build_training_dataset.py from konsman/adu-corpus.

python scripts/data/build_training_dataset.py \
  --output-dir data/training/v1 \
  --push-to-hub konsman/adu-training-v1 \
  --seed 42

Full metadata (exact record counts, weights, caps per split) is stored in dataset_meta.json alongside the JSONL files.

---

Source Licenses

Source	License
arXiv physics (silver)	arXiv non-exclusive license (research use)
Quantum physics lectures (silver)	Source-dependent
SciARG	CC BY 4.0
CSAbstruct	Apache 2.0
PubMed RCT	CC BY 4.0

Silver labels generated by OpenAI gpt-4o via the OpenAI API. Label quality is silver-grade (LLM-generated, confidence-filtered at ≥ 0.80).