readCtrl_lambda / code /finetune-inference /old /inferenceV2_without_context.py

mshahidul

Initial commit of readCtrl code without large models

030876e 7 months ago

5.67 kB

import argparse
import os
import json
import sys
sys.path.append(os.path.abspath('/home/mshahidul/'))
from gpu_selection import _gpu_selection_
# 1. Argparse for path
parser = argparse.ArgumentParser(description="Translation Evaluation")
# parser.add_argument("--out_path", type=str, default="/home/mshahidul/readctrl/generating_data/tik_ache/es_syntheticV3.json", help="Path to the JSON file")
parser.add_argument("--cuda", type=str, default="3", help="CUDA device id, e.g., '0' or '0,1' for multiple GPUs")
parser.add_argument("--model_name", type=str, default="/home/mshahidul/readctrl/finetuned_models/es_synthetic_data_creation_Qwen3_14B_v2", help="Path to the finetuned model")
parser.add_argument("--temperature", type=float, default=0.1, help="Generation temperature")
args = parser.parse_args()
# out_path = args.out_path
model_name = args.model_name
temperature = args.temperature
if args.cuda is not None:
    os.environ["CUDA_VISIBLE_DEVICES"] = args.cuda
    print(f"🎮🎮 Using CUDA device: {args.cuda}")
else:
    _gpu_selection_()

prompts={
"easy":'''
You are an assistant that rewrites Spanish texts to make them very simple and easy to understand.
Your goal is to rewrite the provided input text for younger readers (Fernández Huerta 70–100; grade 5–7).
Use short sentences, simple words, and friendly tone. Avoid technical or complex expressions.
Keep all important factual details, but remove jargon.
Return only the rewritten text without commentary.
''',

'intermediate':'''
You are an assistant specialized in rewriting Spanish texts with medium readability.
Your task is to rewrite the provided input text for general or high‑school‑level readers (Fernández Huerta 50–70; grade 8–12).
Use clear and complete sentences, moderately complex vocabulary, and structured narration.
Retain all relevant medical or factual information, but phrase it in accessible language.
Return only the rewritten text with no explanations.
''',

'hard':'''
You are an assistant that rewrites Spanish medical texts with professional, technical precision.
Rewrite the following input text using specialized, academic terminology and information‑dense phrasing.
The output must target a Fernández Huerta readability index between 0 and 50 (university/professional level).
Use clinical vocabulary, formal register, and detailed description of pathophysiology, procedures, and findings.
Return only the rewritten text.
'''
}

# 2. Output directory and file
path="/home/mshahidul/readctrl/data/testing_data/multiclinsum_test_es.json"
out_dir = "/home/mshahidul/readctrl/results/v2_without_context"
os.makedirs(out_dir, exist_ok=True)
# file_name = os.path.basename(path)
# out_path = os.path.join(out_dir, file_name.replace(".json", "_V2.json"))
# os.makedirs(os.path.dirname(out_dir), exist_ok=True)
if os.path.exists(model_name):
    out_path = out_dir + f"/temp{temperature}_qwen3-14B_finetuned.json"
else:
    out_path = out_dir + f"/temp{temperature}_qwen3-14B_base.json"
# 3. Load already evaluated results if exist
results = []
completed_keys = set()
if os.path.exists(out_path):
    with open(out_path, "r", encoding="utf-8") as f:
        results = json.load(f)
    for r in results:
        completed_keys.add(r["fulltext"])

# 4. Load dataset
with open(path, "r", encoding="utf-8") as f:
    dataset = json.load(f)
dataset=dataset[0:50]  
from unsloth import FastLanguageModel
import torch
# 5. Load model
model, tokenizer = FastLanguageModel.from_pretrained(
    model_name = model_name,
    max_seq_length = 4092,
    load_in_4bit = False,
    load_in_8bit = False,
    full_finetuning = False,
)

import tqdm
for item in tqdm.tqdm(dataset):
    key = item["fulltext"]
    if key in completed_keys:
        continue

    for band in ["easy", "intermediate", "hard"]:
        prompt = prompts[band]+'\n\n'+"Input text:\n"+item['fulltext']

        # messages = [{"role": "user", "content": prompt+"\n"}]
        messages = [
                {"role": "system", "content": prompts[band].strip()},
                {"role": "user", "content": "Input text:\n" + item["fulltext"].strip()}
                ]
        text = tokenizer.apply_chat_template(
            messages,
            tokenize=False,
            add_generation_prompt=True,
            enable_thinking=False,
        )
        # input_ids = tokenizer(item["fulltext"], return_tensors="pt").input_ids
        # input_len = input_ids.shape[1]
        inputs = tokenizer(text, return_tensors="pt").to("cuda")
        input_len = inputs.input_ids.shape[1]
        # Define proportional multipliers for each readability level
        length_factors = {"easy": 0.5, "intermediate": 0.8, "hard": 1.1}

        # Compute adaptive max_new_tokens
        max_new_tokens = int(min(1200, max(150, input_len * length_factors[band])))
        output_ids = model.generate(
            **inputs,
            max_new_tokens=max_new_tokens,
            temperature=temperature,
            top_p=0.9,
            top_k=45,
        )
        output_text = tokenizer.decode(output_ids[0], skip_special_tokens=True)
        #answer = output_text.split("</think>")[1].strip()

        results.append({
            "fulltext": item["fulltext"],
            "band": band,
            "lang": "es",
            "synthetic_summary": output_text,
        })
        completed_keys.add(key)
        # Save every 10 results
        if len(results) % 3 == 0:
            with open(out_path, "w", encoding="utf-8") as f:
                json.dump(results, f, ensure_ascii=False, indent=2)

# 7. Final save
with open(out_path, "w", encoding="utf-8") as f:
    json.dump(results, f, ensure_ascii=False, indent=2)

	import argparse
	import os
	import json
	import sys
	sys.path.append(os.path.abspath('/home/mshahidul/'))
	from gpu_selection import _gpu_selection_
	# 1. Argparse for path
	parser = argparse.ArgumentParser(description="Translation Evaluation")
	# parser.add_argument("--out_path", type=str, default="/home/mshahidul/readctrl/generating_data/tik_ache/es_syntheticV3.json", help="Path to the JSON file")
	parser.add_argument("--cuda", type=str, default="3", help="CUDA device id, e.g., '0' or '0,1' for multiple GPUs")
	parser.add_argument("--model_name", type=str, default="/home/mshahidul/readctrl/finetuned_models/es_synthetic_data_creation_Qwen3_14B_v2", help="Path to the finetuned model")
	parser.add_argument("--temperature", type=float, default=0.1, help="Generation temperature")
	args = parser.parse_args()
	# out_path = args.out_path
	model_name = args.model_name
	temperature = args.temperature
	if args.cuda is not None:
	os.environ["CUDA_VISIBLE_DEVICES"] = args.cuda
	print(f"🎮🎮 Using CUDA device: {args.cuda}")
	else:
	_gpu_selection_()

	prompts={
	"easy":'''
	You are an assistant that rewrites Spanish texts to make them very simple and easy to understand.
	Your goal is to rewrite the provided input text for younger readers (Fernández Huerta 70–100; grade 5–7).
	Use short sentences, simple words, and friendly tone. Avoid technical or complex expressions.
	Keep all important factual details, but remove jargon.
	Return only the rewritten text without commentary.
	''',

	'intermediate':'''
	You are an assistant specialized in rewriting Spanish texts with medium readability.
	Your task is to rewrite the provided input text for general or high‑school‑level readers (Fernández Huerta 50–70; grade 8–12).
	Use clear and complete sentences, moderately complex vocabulary, and structured narration.
	Retain all relevant medical or factual information, but phrase it in accessible language.
	Return only the rewritten text with no explanations.
	''',

	'hard':'''
	You are an assistant that rewrites Spanish medical texts with professional, technical precision.
	Rewrite the following input text using specialized, academic terminology and information‑dense phrasing.
	The output must target a Fernández Huerta readability index between 0 and 50 (university/professional level).
	Use clinical vocabulary, formal register, and detailed description of pathophysiology, procedures, and findings.
	Return only the rewritten text.
	'''
	}

	# 2. Output directory and file
	path="/home/mshahidul/readctrl/data/testing_data/multiclinsum_test_es.json"
	out_dir = "/home/mshahidul/readctrl/results/v2_without_context"
	os.makedirs(out_dir, exist_ok=True)
	# file_name = os.path.basename(path)
	# out_path = os.path.join(out_dir, file_name.replace(".json", "_V2.json"))
	# os.makedirs(os.path.dirname(out_dir), exist_ok=True)
	if os.path.exists(model_name):
	out_path = out_dir + f"/temp{temperature}_qwen3-14B_finetuned.json"
	else:
	out_path = out_dir + f"/temp{temperature}_qwen3-14B_base.json"
	# 3. Load already evaluated results if exist
	results = []
	completed_keys = set()
	if os.path.exists(out_path):
	with open(out_path, "r", encoding="utf-8") as f:
	results = json.load(f)
	for r in results:
	completed_keys.add(r["fulltext"])

	# 4. Load dataset
	with open(path, "r", encoding="utf-8") as f:
	dataset = json.load(f)
	dataset=dataset[0:50]
	from unsloth import FastLanguageModel
	import torch
	# 5. Load model
	model, tokenizer = FastLanguageModel.from_pretrained(
	model_name = model_name,
	max_seq_length = 4092,
	load_in_4bit = False,
	load_in_8bit = False,
	full_finetuning = False,
	)

	import tqdm
	for item in tqdm.tqdm(dataset):
	key = item["fulltext"]
	if key in completed_keys:
	continue

	for band in ["easy", "intermediate", "hard"]:
	prompt = prompts[band]+'\n\n'+"Input text:\n"+item['fulltext']

	# messages = [{"role": "user", "content": prompt+"\n"}]
	messages = [
	{"role": "system", "content": prompts[band].strip()},
	{"role": "user", "content": "Input text:\n" + item["fulltext"].strip()}
	]
	text = tokenizer.apply_chat_template(
	messages,
	tokenize=False,
	add_generation_prompt=True,
	enable_thinking=False,
	)
	# input_ids = tokenizer(item["fulltext"], return_tensors="pt").input_ids
	# input_len = input_ids.shape[1]
	inputs = tokenizer(text, return_tensors="pt").to("cuda")
	input_len = inputs.input_ids.shape[1]
	# Define proportional multipliers for each readability level
	length_factors = {"easy": 0.5, "intermediate": 0.8, "hard": 1.1}

	# Compute adaptive max_new_tokens
	max_new_tokens = int(min(1200, max(150, input_len * length_factors[band])))
	output_ids = model.generate(
	**inputs,
	max_new_tokens=max_new_tokens,
	temperature=temperature,
	top_p=0.9,
	top_k=45,
	)
	output_text = tokenizer.decode(output_ids[0], skip_special_tokens=True)
	#answer = output_text.split("</think>")[1].strip()

	results.append({
	"fulltext": item["fulltext"],
	"band": band,
	"lang": "es",
	"synthetic_summary": output_text,
	})
	completed_keys.add(key)
	# Save every 10 results
	if len(results) % 3 == 0:
	with open(out_path, "w", encoding="utf-8") as f:
	json.dump(results, f, ensure_ascii=False, indent=2)

	# 7. Final save
	with open(out_path, "w", encoding="utf-8") as f:
	json.dump(results, f, ensure_ascii=False, indent=2)