Extracts softConstraints JSON from free-text employee crew notes and manager limitation instructions. Output schema includes: dailyTimeRestrictions, weeklyFrequencyLimits, consecutiveShiftLimits, recurringTimeOffPatterns, crossDayDependencies, advanceNoticeRequired, crewSizeRestrictions, leadershipRestrictions, jobTypeRestrictions, clientScheduleRestrictions, vehicleRestrictions, interpersonalConflicts.

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch, json

model = AutoModelForCausalLM.from_pretrained("loitranyuki/aischeduler-llm-v4", torch_dtype=torch.float16, device_map="auto")
tokenizer = AutoTokenizer.from_pretrained("loitranyuki/aischeduler-llm-v4")

system_prompt = "..."  # see prompts/employee_constraint_extraction.txt
user_text = "No evenings. Max 3 doubles per week."

messages = [
    {"role": "system", "content": system_prompt},
    {"role": "user", "content": user_text},
]
tokens = tokenizer.apply_chat_template(messages, return_tensors="pt").to("cuda")
out = model.generate(tokens, max_new_tokens=512, temperature=0.1)
raw = tokenizer.decode(out[0][tokens.shape[-1]:], skip_special_tokens=True)
constraints = json.loads(raw)

Training

Framework: Unsloth + PEFT (LoRA rank=16, alpha=32)
Quantization: 4-bit QLoRA during training, merged to bfloat16
Target modules: q_proj, k_proj, v_proj, o_proj