southern-lev-fat
advanced citation analysis in judicial rulings
2 min readupdated
The southern-lev-fat model is an updated version of a tool for classifying legal citations in judicial decisions. It is fine-tuned on ModernBERT-large (an encoder-based architecture) with an expanded input context of up to 1,500 bytes per citation occurrence, allowing it to capture the discussion context more accurately. The model classifies citation sentiment across 10 labels, including negative treatments such as overruling or limiting precedent. The development is geared toward legal analysis automation and jurisprudence researchers.
Key features
- Wide input context: captures 1,500 bytes of text per citation occurrence with a proportion of 30% before and 70% after the mention.
- 10 label categories: determines how a citation is applied (for example,
follows,distinguishes,overrules). - Agent-assisted training: new training pairs are labeled using Claude Fable 5.2 agents based on U.S. federal court materials.
- Improved metrics: shows growth in AUROC (area under the ROC curve) and better detects negative treatments compared to the baseline version.
What you can do
- Jurisprudence analysis: automate the verification of how lower courts apply or criticize precedent decisions.
- Legal document auditing: use the model in legal text processing pipelines to find hidden contradictions in citations.
- Script integration: use a custom context window generation script to process large arrays of judicial opinions.
Example
Example of a function for generating context windows from citation text:
def windows(text, hits, width=1500, before=0.3, n=3):
b = text.encode("utf8")
chosen = []
order = list(dict.fromkeys([0, len(hits) // 2, len(hits) - 1, *range(len(hits))]))
for i in order:
h = hits[i]
centre = (h["start"] + h["end"]) // 2
lo = max(0, centre - int(width * before))
hi = min(len(b), lo + width)
if any(lo < y and x < hi for x, y in chosen):
continue
chosen.append((lo, hi))
if len(chosen) == n:
break
return chosenRead more
DetailsBase model: ferguscooney/southern-lev (ModernBERT-large)
Parameters: 0.4B
Training: three epochs, batch size 2 with gradient accumulation 4, fp32 on Apple MPS
Metrics: AUROC 0.903 on an independent Demir & Canbaz dataset










Comments