qwen2.5-1.5b-earnings-risk-extractor

Earnings-Call Risk Extractor (Qwen2.5-1.5B-Instruct, QLoRA)

Fine-tuned to extract a structured risk register from earnings-call and MD&A commentary. Built for the CDAZZDEV Senior ML Engineer technical assessment (Task 2).

Task

Input: a passage of management commentary plus the sector. Output: JSON — a list of disclosed risks, each with a category from a closed 10-item taxonomy, a severity, a time horizon, and an evidence span quoted verbatim from the passage.

Training

Usage

The model was trained and evaluated with this exact system prompt; use it unchanged.

from transformers import AutoModelForCausalLM, AutoTokenizer

SYSTEM_PROMPT = '''You extract a structured risk register from earnings-call commentary.

Return ONLY a JSON object: {"risks": [{"risk_category", "description", "severity", "time_horizon", "evidence_span"}], "overall_risk_tone"}.

risk_category must be one of: SUPPLY_CHAIN, DEMAND, COMPETITION, REGULATORY, FX_MACRO, LIQUIDITY, TALENT, TECHNOLOGY, CYBER_SECURITY, LEGAL_LITIGATION.
evidence_span must be copied verbatim from the passage.
severity: high | medium | low. time_horizon: near_term | medium_term | long_term.
overall_risk_tone: elevated | moderate | contained.'''

tok = AutoTokenizer.from_pretrained("dineth9d/qwen2.5-1.5b-earnings-risk-extractor")
model = AutoModelForCausalLM.from_pretrained("dineth9d/qwen2.5-1.5b-earnings-risk-extractor", device_map="auto")

messages = [
    {"role": "system", "content": SYSTEM_PROMPT},
    {"role": "user", "content": "Sector: semiconductors\n\nPASSAGE:\n\"\"\"\n<management commentary>\n\"\"\"\n\nExtract the risk register."},
]
enc = tok.apply_chat_template(messages, add_generation_prompt=True, return_dict=True, return_tensors="pt").to(model.device)
out = model.generate(**enc, max_new_tokens=640, do_sample=False)
print(tok.decode(out[0][enc["input_ids"].shape[-1]:], skip_special_tokens=True))

Limitations

Trained on synthetic passages generated by a teacher model, not on real filings. It inherits the teacher's stylistic register and its blind spots, and it has not been validated against SEC documents. Evidence spans are grounded by construction but should still be verified against source before any downstream use. Not investment advice.