fakechamp24/legallens-bart-base

LegalLens BART-base (3 epochs)

facebook/bart-base fine-tuned for chunk-level summarization of Indian court judgements, as the summarizer in the LegalLens pipeline (section detection → chunking → per-chunk summary → Lawyer / Citizen modes → Cite & Verify).

Training data

Section-aligned pseudo-targets derived from percins/IN-ABS: each judgement is chunked by section, and each sentence of its reference summary is aligned to the chunk it came from with a cross-encoder. Inputs are chunk_text, targets are aligned_target.

• 31,029 training / 4,198 validation records (5,346 / 712 judgements; official IN-ABS splits)
• 17.9% of reference-summary sentences could not be aligned and are not used as targets
• Argument sections are thin (4.6% of records)

Training

3 epochs, lr 3e-5, effective batch 32 (4 × 8 accumulation), warmup 6%, weight decay 0.01, label smoothing 0.1, bf16, max input 1024 / target 448 tokens. Best checkpoint = last epoch. Chunk-level validation ROUGE-1/2/L: 38.74 / 22.37 / 34.72.

Evaluation (document level)

First 100 IN-ABS test judgements, full LegalLens pipeline, scored against the published summary:

Model	Output	ROUGE-1	ROUGE-2	ROUGE-L	BERTScore F1
bart-base zero-shot	Lawyer Mode	50.93	24.96	44.65	83.51
this model	Lawyer Mode	55.15	30.75	49.64	85.04
bart-base zero-shot	all chunk summaries	52.86	30.05	45.11	83.67
this model	all chunk summaries	58.80	34.04	51.07	85.13

Usage

from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
tok = AutoTokenizer.from_pretrained("fakechamp24/legallens-bart-base")
model = AutoModelForSeq2SeqLM.from_pretrained("fakechamp24/legallens-bart-base")
inputs = tok(chunk_text, max_length=1024, truncation=True, return_tensors="pt")
summary = tok.decode(model.generate(**inputs, num_beams=4, max_length=250)[0], skip_special_tokens=True)

In the LegalLens repo: LEGAL_BART_MODEL=fakechamp24/legallens-bart-base python cli.py summarize <file>.

Limitations

• Trained on Indian Supreme Court judgements in English; other jurisdictions transfer poorly.
• Targets are cross-encoder-aligned pseudo-targets, not human per-section summaries.
• Generated summaries are largely extractive and can garble long lists of case citations; use the pipeline's Cite & Verify output to flag unsupported sentences.
• Not legal advice.

Licence

percins/IN-ABS declares no licence, so the licence of this derived model is undetermined. Keep the repository private until the dataset authors' terms are confirmed.