ihlee1/navio-flow-v1.0

navio-flow v1.0 — Maritime Email Classifier

LoRA adapter fine-tuned from Qwen/Qwen3-4B-Instruct-2507 for 4-class multi-label maritime chartering email classification.

Categories

• vessel — Owner offers ship position (Open Tonnage)
• cargo — Charterer seeks ship for cargo (Cargo Inquiry)
• tct — Time Charter Trip (single voyage with daily hire rate)
• other — S&P, market reports, ops, admin

Multi-label predictions allowed (e.g. ["cargo", "tct"]).

Test Set Performance (n=301)

Metric	Value
Macro F1	0.7639
Set accuracy	0.7608

Per-class

Class	Precision	Recall	F1
vessel	0.887	0.683	0.772
cargo	0.984	0.775	0.867
tct	0.714	0.625	0.667
other	0.621	0.947	0.750

Training Details

• Framework: Unsloth 2026.4.8
• Base model: Qwen/Qwen3-4B-Instruct-2507
• Trainable params: 66,060,288 / 4,088,528,384 (1.62%)
• Iters: 475
• Wall-clock training time: 90.8 min
• Final train loss: 0.7736
• Hardware: NVIDIA A100-SXM4-40GB
• Snapshot hash: b817d2b49800abae

Hyperparameters

• LoRA rank: 32, alpha: 64, dropout: 0.05
• Target modules: q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj
• Learning rate: 1e-4 (cosine schedule, 10% warmup)
• Effective batch size: 32 (per_device 2 × grad_accum 16)
• Epochs: 5
• Max sequence length: 2048
• 4-bit QLoRA via bitsandbytes

Usage

from unsloth import FastLanguageModel

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name="ihlee1/navio-flow-v1.0",
    max_seq_length=2048,
    load_in_4bit=True,
)
FastLanguageModel.for_inference(model)

# Use the system prompt from the snapshot
system_prompt = open("_system_prompt.txt").read()  # 580-token dense ruleset

messages = [
    {"role": "system", "content": system_prompt},
    {"role": "user", "content": "Subject: ...\n\nDELY Busan, REDELY Spore..."},
]
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(prompt, return_tensors="pt").to("cuda")
out = model.generate(**inputs, max_new_tokens=40)
print(tokenizer.decode(out[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))
# Expected output: {"labels": ["vessel"]}

Limitations

• Macro-F1 0.764 — below the V1.0 target of 0.93. Issues:
  • tct minority class (test n=19) — F1 0.67
  • Long emails truncated at max_seq_len=2048 (some originals 3000-3800 tokens)
  • Multi-label cases sparse in test set (n=5)
• Trained on emails from one specific brokerage (~3000 examples). Generalization to other brokerages untested.

License

Apache 2.0 (inherits from Qwen3-4B-Instruct-2507).