File size: 5,035 Bytes
6e644fe |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 |
import json
import re
from collections import defaultdict
from .data_utils import load_article_text
from .prompt_utils import build_prompt
from .metrics import exact_match,f1_score,hamming_score,lca_score,clean_model_answer
#Model evaluation
def evaluate(llm,tokenizer,sampling_params,questions,article_path,with_think=False):
emr_total=0
f1_total=0
lca_total=0
hamming_total=0
count=0
invalid_count=0
extracted_from_invalid=0
label_scores=defaultdict(lambda: {"emr":[],"f1":[],"lca":[],"hamming":[]})
results_list=[]
for q in questions:
article=load_article_text(q["id_article"],article_path)
if not article:
continue
prompt=build_prompt(article,q,tokenizer)
if with_think:
prompt=prompt
else:
prompt="/no_think\n"+prompt
outputs=llm.generate(prompts=[prompt],sampling_params=sampling_params)
model_raw_output=outputs[0].outputs[0].text.strip()
if with_think:
justification_match=re.search(r"<think>\s*(.*?)\s*</think>",model_raw_output, re.DOTALL | re.IGNORECASE)
justification_model=justification_match.group(1).strip() if justification_match else ""
else:
justification_model=""
print(f"\nQuestion {count+1}/{len(questions)} traitée",flush=True)
print(f"\nRaw model output: {model_raw_output}",flush=True)
is_valid_format=bool(re.match(r"^([a-eA-E](?:[,\s]+[a-eA-E])*)$",model_raw_output.strip()))
if not is_valid_format:
invalid_count+=1
cleaned=clean_model_answer(model_raw_output)
if cleaned:
extracted_from_invalid+=1
model_answer=clean_model_answer(model_raw_output)
model_answer=model_answer.lower().strip().replace(",","").replace(" ","")
gold_answer="".join(q["correct_answers"]).lower()
emr=exact_match(model_answer,gold_answer)
f1=f1_score(model_answer,gold_answer)
hamming=hamming_score(model_answer,gold_answer)
lca=lca_score(model_answer,gold_answer)
emr_total+=int(emr)
f1_total+=f1
hamming_total+=hamming
lca_total+=lca
count+=1
for label in q["labels"]:
label=str(label)
label_scores[label]["emr"].append(int(emr))
label_scores[label]["f1"].append(f1)
label_scores[label]["lca"].append(lca)
label_scores[label]["hamming"].append(hamming)
print(f"\nQ: {q['question']}\nPredicted: {model_answer} | Gold: {gold_answer} | EMR: {emr} | F1: {f1:.2f} | Hamming: {hamming:.2f} | LCA-score: {lca:.2f}", flush=True)
#Add results per question
results_list.append({
"question_id": q.get("id"),
"article_id": q.get("id_article"),
"article_date": q.get("article_date"),
"exam_date": q.get("date_exam"),
"question": q["question"],
"raw_output": model_raw_output,
"predicted_answers": list(model_answer.lower()),
"gold_answers": q["correct_answers"],
"justification_model": justification_model,
"justification_human": q.get("justification"),
"emr": round(emr,2),
"f1": round(f1,2),
"hamming": round(hamming,2),
"lca": lca
})
#Results summary
summary={
"summary_global":{
"total_evaluated": count,
"emr_avg": round(emr_total/count,2) if count>0 else 0,
"f1_avg": round(f1_total/count,2) if count>0 else 0,
"hamming_avg": round(hamming_total/count,2) if count>0 else 0,
"lca_avg": round(lca_total/count,2) if count>0 else 0,
"malformatted_percentage": round((invalid_count/count)*100,2) if count>0 else 0,
"extraction_rate": round((extracted_from_invalid/invalid_count*100),2) if invalid_count>0 else 0
},
"summary_by_label": {}
}
for label in sorted(label_scores.keys()):
label_emr_avg=sum(label_scores[label]["emr"])/len(label_scores[label]["emr"])
label_f1_avg=sum(label_scores[label]["f1"])/len(label_scores[label]["f1"])
label_lca_avg=sum(label_scores[label]["lca"])/len(label_scores[label]["lca"])
label_hamming_avg=sum(label_scores[label]["hamming"])/len(label_scores[label]["hamming"])
summary["summary_by_label"][label] = {
"emr_avg": round(label_emr_avg,2),
"f1_avg": round(label_f1_avg,2),
"hamming_avg": round(label_hamming_avg,2),
"lca_avg": round(label_lca_avg,2)
}
#Save results to a JSON file
with open("evaluation_results.json","w",encoding="utf-8") as f:
json.dump({
"results": results_list,
"summary": summary
}, f,indent=2,ensure_ascii=False)
print("\n=== Results saved in 'evaluation_results.json' ===",flush=True) |