Downloads · 30 days
0
davide-taglialatela/disaster-interpretation-qwen3.5-9b
disaster-interpretation-qwen3.5-9b is a text generation model from davide-taglialatela. Use it when you need the model to write or continue text. It is set up for peft. The card lists the license as cc-by-4.0.
disaster-interpretation-qwen3.5-9b is a LoRA adapter for unsloth/Qwen3.5-9B that classifies paragraphs of U.S. newspaper coverage of natural disasters, answering in JSON. It was built for the paper What a Disaster Mea…
Downloads · 30 days
0
Access
Public
Updated Oct 26, 2026
Repo size
136 MB
Likes
0
Public
Click a slice to open those files.
.safetensors116 MB · 85%
From the Hugging Face model README
disaster-interpretation-qwen3.5-9b is a LoRA adapter for unsloth/Qwen3.5-9B that classifies paragraphs of U.S. newspaper coverage of natural disasters, answering in JSON. It was built for the paper What a Disaster Means: Attention, Interpretation, and the Demand for Green Assets (Taglialatela, 2026).
It is the most accurate of the models trained for the paper. There, disaster-interpretation-deberta-v3-large scores every paragraph and this model re-scores the paragraphs on which the encoder's confidence on the climate link or on skepticism is below 0.67 (about 29,000 of 15.5 million); its label is then used.
Each paragraph is classified on its own, without the headline or the surrounding text, on 12 heads:
| Head | Description | Classes |
|---|---|---|
gate | Main subject of the paragraph | disaster_specific, disaster_general, climate_no_disaster, peripheral, irrelevant |
peripheral_descriptor | Type of policy or market content (peripheral paragraphs only) | disaster-policy, climate-policy, green, insurance, prep |
specific_event_present | A concrete, located and dated hazard event is present | 0, 1 |
link | Link between the disaster and climate change | none, passing, central |
literal_stance | Stance on the link | affirms, neutral, denies |
voice | Whose stance it is | own, quoted |
link_direction | Direction of the link, net of voice | affirms, neutral_contested, denies, na |
minimize | The link is played down | 0, 1 |
anti_climate_sarcasm | Sarcasm aimed at climate concern | 0, 1 |
denial_coarse | Type of denial (denying paragraphs only) | a_science, b_notbad, c_solutions, d_notus |
human | Affected people in the paragraph | none, present, central |
personal_story | A named person's own story | 0, 1 |
The gate head is read first. The link, event, human and story heads apply only when the gate is disaster_specific or disaster_general. The stance heads apply when the gate is climate_no_disaster, when a disaster paragraph has a passing or central link, or when a peripheral paragraph makes a clear climate claim. denial_coarse applies only when the stance is denies; peripheral_descriptor only when the gate is peripheral. Outputs of heads that do not apply should be discarded.
The model is prompted in four stages (module_prompts.json: P_GATE, P_DISASTER_LINK, P_STANCE, P_STORY_HUMAN). Each stage receives its codebook as the system message and the paragraph as the user message, preceded by the earlier decisions. Greedy decoding, at most 160 new tokens, thinking disabled. Requires a GPU, Unsloth and transformers 4.57 or newer.
import json, re
from huggingface_hub import snapshot_download
from unsloth import FastLanguageModel
run = snapshot_download("davide-taglialatela/disaster-interpretation-qwen3.5-9b")
model, tok = FastLanguageModel.from_pretrained(f"{run}/adapter", max_seq_length=8192, load_in_4bit=True)
FastLanguageModel.for_inference(model)
prompts = json.load(open(f"{run}/module_prompts.json"))
def ask(module, block, prior=""):
user = (prior + "\n\n" if prior else "") + "BLOCK:\n" + block.strip()
msgs = [{"role": "system", "content": [{"type": "text", "text": prompts[module]}]},
{"role": "user", "content": [{"type": "text", "text": user}]}]
ids = tok.apply_chat_template(msgs, tokenize=True, add_generation_prompt=True,
return_tensors="pt", enable_thinking=False).to(model.device)
out = model.generate(input_ids=ids, max_new_tokens=160, do_sample=False)
txt = tok.decode(out[0][ids.shape[1]:], skip_special_tokens=True)
m = re.search(r"\{.*?\}", txt, re.S)
return json.loads(m.group(0)) if m else {}
def classify(block):
r = ask("P_GATE", block); gate = r.get("gate")
prior = f"PRIOR STAGE RESULT (already decided — take as fixed): gate={gate}"
if gate in ("disaster_specific", "disaster_general"):
r.update(ask("P_DISASTER_LINK", block, prior))
r.update(ask("P_STORY_HUMAN", block, prior))
link = r.get("link")
if gate in ("climate_no_disaster", "peripheral") or link in ("passing", "central"):
r.update(ask("P_STANCE", block, prior + (f", link={link}" if link else "")))
return r
The task knowledge is in the adapter and the prompts together. The base model without the adapter, or the adapter merged into another base, does not perform the task.
The model was trained on English-language U.S. newspaper paragraphs about natural hazards and climate change, and has not been evaluated on other text. The training labels come from a language model, not from human annotators; agreement with the author's hand labels is reported below, an independent annotator study is not yet available. The model was trained on Nexis Uni and Wall Street Journal text and applied, in the paper, to ProQuest TDM Studio newspapers. Predictions are for one paragraph at a time; the paper aggregates them to the article (an article is linked if any paragraph is).
The training set is 5,228 paragraphs from 3,003 U.S. newspaper articles (Nexis Uni local newspapers and the Wall Street Journal, 1988–2026, mostly after 2010). Paragraphs likely to contain rare classes (denial, sarcasm, central links) were oversampled.
Labels were produced by GPT-5.5 (gpt-5.5-2026-04-23) under a written codebook (version 4.3.4). The codebook was developed against about 300 paragraphs labelled by the author. 270 paragraphs of the test set also carry the author's hand labels; they are used for evaluation only.
| Split | Paragraphs | Articles |
|---|---|---|
| Train | 5,228 | 3,003 |
| Validation | 696 | 387 |
| Test, random | 639 | 366 |
| Test, hard (hand-labelled articles) | 760 | 245 |
Splits are by article. No article text is distributed in this repository.
QLoRA with Unsloth on the 4-bit base model: rank 16, alpha 32, no dropout, on the q, k, v, o, gate, up and down projections. Supervised fine-tuning on the 5,228 training paragraphs with the loss on the answer only, one epoch, learning rate 0.0002, batch size 4 with 4 accumulation steps, 8-bit AdamW, linear schedule with 5% warm-up, sequences up to 8192 tokens, bf16, seed 42. Every output on the evaluation splits parsed as JSON.
Positive-class F1 of each candidate model against the reference labels on held-out paragraphs. disaster: gate is disaster_specific or disaster_general; specific: specific_event_present; link: link is passing or central (the paper's main variable); skeptical: direction denies, or minimize, or sarcasm; human: human is present or central; flip: share of paragraphs whose direction the model reverses. The gpt rows are general-purpose models prompted with the same codebook.
| Model | disaster | specific | link | skeptical | human | flip |
|---|---|---|---|---|---|---|
| Qwen3.5-9B-staged | .952 | .960 | .947 | .911 | .962 | .015 |
| DeBERTa-v3-large | .905 | .932 | .938 | .880 | .909 | .015 |
| ModernBERT-large | .891 | .900 | .927 | .867 | .878 | .034 |
| gpt-5.4-mini | .877 | .917 | .906 | .818 | .954 | .053 |
| gpt-4.1-mini | .893 | .895 | .899 | .774 | .893 | .075 |
| gpt-4o-mini | .745 | .931 | .739 | .804 | .911 | .063 |
Against the author's hand labels on the hard test set:
| Model | disaster | link | skeptical | flip |
|---|---|---|---|---|
| Qwen3.5-9B-staged | .888 | .876 | .918 | .009 |
| DeBERTa-v3-large | .876 | .868 | .871 | .027 |
| ModernBERT-large | .817 | .850 | .885 | .045 |
| gpt-5.4-mini | .804 | .768 | .830 | .029 |
| gpt-4.1-mini | .869 | .846 | .800 | .059 |
| gpt-4o-mini | .667 | .638 | .844 | .077 |
Per-head macro-F1 and confusion matrices are in the repository's evaluation files.
Paper: latest version. Author website: davide-taglialatela.github.io.
@unpublished{taglialatela2026disaster,
title = {What a Disaster Means: Attention, Interpretation, and the Demand for Green Assets},
author = {Taglialatela, Davide},
year = {2026},
note = {Job market paper, Harvard University},
url = {https://davide-taglialatela.github.io/files/JMP_Davide_Taglialatela.pdf}
}
The paper applies the model to 1.49 million disaster articles from 527 U.S. news outlets and builds a county-level measure of how much coverage each disaster received and how much of it interpreted the disaster as climate change. That measure is not yet public; it will be released as a Hugging Face dataset with the paper.