kira_train.py

#!/usr/bin/env python3
import os
os.environ["HF_HOME"] = "/workspace/hf_cache"
os.environ["TRANSFORMERS_OFFLINE"] = "1"
os.environ["HF_DATASETS_OFFLINE"] = "1"
os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "expandable_segments:True"

import unsloth
from unsloth import FastLanguageModel, is_bfloat16_supported

import json
import shutil
import torch
from datasets import Dataset
from trl import SFTTrainer, SFTConfig

MODEL = "/workspace/hf_cache/hub/models--unsloth--gemma-4-26B-A4B-it/snapshots/cd98c13581a9d4ad061cb85d983232ca4edb1343"
DATA_DIR = "/workspace/kira_gemma4_training/data/ft"
OUTPUT_DIR = "/workspace/kira_gemma4_training/kira_out"
FINAL_DIR = "/workspace/kira_gemma4_training/kira_adapter"
LOG_DIR = "/workspace/kira_gemma4_training/kira_logs"
TOKEN_CACHE = "/workspace/kira_gemma4_training/tokenized_cache"

MAX_SEQ = 1024
NUM_PROC = 16

SYSTEM_PROMPT = (
    "You are Kira, a contract review assistant. Given a clause and its surrounding context, "
    "identify the issue type, assess severity, explain what is wrong, and describe the worst-case "
    "consequence. Respond in strict JSON with keys: issue_type, severity, severity_rationale, "
    "what_is_wrong, worst_case, evidence_span."
)

def load_jsonl(path):
    with open(path, encoding="utf-8") as f:
        return [json.loads(line) for line in f if line.strip()]

def build_user_msg(ex):
    parts = []
    if ex.get("clause_type"):
        parts.append(f"Clause type: {ex['clause_type']}")
    if ex.get("left_context"):
        parts.append(f"Left context:\n{ex['left_context']}")
    parts.append(f"Clause:\n{ex.get('clause_text','')}")
    if ex.get("right_context"):
        parts.append(f"Right context:\n{ex['right_context']}")
    if ex.get("cuad_question"):
        parts.append(f"Question: {ex['cuad_question']}")
    return "\n\n".join(parts)

def build_assistant_msg(ex):
    out = {
        "issue_type": ex.get("ds_issue_type"),
        "severity": ex.get("ds_severity"),
        "severity_rationale": ex.get("ds_severity_rationale"),
        "what_is_wrong": ex.get("ds_what_is_wrong"),
        "worst_case": ex.get("ds_worst_case"),
        "evidence_span": ex.get("ds_evidence_span"),
    }
    return json.dumps(out, ensure_ascii=False)

print("Loading model...", flush=True)
model, tokenizer = FastLanguageModel.from_pretrained(
    model_name=MODEL,
    max_seq_length=MAX_SEQ,
    dtype=None,
    load_in_4bit=True,
    full_finetuning=False,
)

print("Applying LoRA...", flush=True)
model = FastLanguageModel.get_peft_model(
    model,
    r=16,
    lora_alpha=32,
    lora_dropout=0.0,
    bias="none",
    target_modules=[
        "q_proj", "k_proj", "v_proj", "o_proj",
        "gate_proj", "up_proj", "down_proj",
    ],
    use_gradient_checkpointing="unsloth",
    random_state=3407,
)

def format_sample(ex):
    if "messages" in ex:
        msgs = ex["messages"]
    else:
        msgs = [
            {"role": "system", "content": SYSTEM_PROMPT},
            {"role": "user", "content": build_user_msg(ex)},
            {"role": "assistant", "content": build_assistant_msg(ex)},
        ]
    return tokenizer.apply_chat_template(
        msgs,
        tokenize=False,
        add_generation_prompt=False,
    )

print("Loading data...", flush=True)
if os.path.isdir(TOKEN_CACHE + "/train") and os.path.isdir(TOKEN_CACHE + "/val"):
    train_data = Dataset.load_from_disk(TOKEN_CACHE + "/train")
    val_data = Dataset.load_from_disk(TOKEN_CACHE + "/val")
    print("Loaded token cache.", flush=True)
else:
    train_data = Dataset.from_list(
        [{"text": format_sample(x)} for x in load_jsonl(DATA_DIR + "/train.jsonl")]
    )
    val_data = Dataset.from_list(
        [{"text": format_sample(x)} for x in load_jsonl(DATA_DIR + "/val.jsonl")]
    )
    os.makedirs(TOKEN_CACHE, exist_ok=True)
    train_data.save_to_disk(TOKEN_CACHE + "/train")
    val_data.save_to_disk(TOKEN_CACHE + "/val")

print(f"Train: {len(train_data)} Val: {len(val_data)}", flush=True)

os.makedirs(OUTPUT_DIR, exist_ok=True)
os.makedirs(LOG_DIR, exist_ok=True)

trainer = SFTTrainer(
    model=model,
    tokenizer=tokenizer,
    train_dataset=train_data,
    eval_dataset=val_data,
    args=SFTConfig(
        output_dir=OUTPUT_DIR,
        num_train_epochs=1,
        per_device_train_batch_size=6,
        gradient_accumulation_steps=8,
        warmup_ratio=0.03,
        learning_rate=2e-4,
        lr_scheduler_type="cosine",
        bf16=is_bfloat16_supported(),
        fp16=not is_bfloat16_supported(),
        optim="adamw_8bit",
        weight_decay=0.01,
        logging_steps=10,
        eval_strategy="steps",
        eval_steps=500,
        save_strategy="steps",
        save_steps=500,
        save_total_limit=3,
        load_best_model_at_end=True,
        metric_for_best_model="eval_loss",
        greater_is_better=False,
        logging_dir=LOG_DIR,
        dataset_text_field="text",
        max_seq_length=MAX_SEQ,
        packing=False,
        seed=3407,
        dataset_num_proc=NUM_PROC,
        report_to=[],
        run_name="kira-gemma4-26b-a4b-qlora",
    ),
)

print("Starting training...", flush=True)
trainer.train()

print("Saving adapter...", flush=True)
model.save_pretrained(OUTPUT_DIR)
tokenizer.save_pretrained(OUTPUT_DIR)
os.makedirs(FINAL_DIR, exist_ok=True)
shutil.copytree(OUTPUT_DIR, FINAL_DIR, dirs_exist_ok=True)
print(f"Done. Adapter at {FINAL_DIR}", flush=True)