CoCo-Decision-DS-4B

Version 1.0.0 (2026-10-08) · Git tag v1.0.0 in this repository

CoCo-Decision-DS-4B is a 4B-parameter decision model for disaster safety (DS) response. It is a domain-adapted version of CoCo-Decision-4B-Ko. Given a situation and a typed question, it returns a probability distribution instead of free text:

Question type	Output
noul	probability that a yes/no statement holds
choice	probability for each of the listed options
score	probability for each point on an ordinal scale

It is a LoRA adapter on Qwen/Qwen3.5-4B (revision 851bf6e8). Probabilities are read from the option-label token logits of a single forward pass, so one decision costs one prefill and no generation.

Not affiliated with or endorsed by TypeSafe AI. "System One" and "Jev" are names used by TypeSafe AI; this model independently implements a compatible decision interface.

Intended use

• Decision support for an incident commander in facility disaster response: classifying incoming field reports and sensor events, proposing the next procedure step, and flagging when a human confirmation is required
• Places where a threshold on a probability is more useful than a generated answer (for example, propose an action above 0.9, ask the operator below it)

Every proposed action must be approved by a responsible person. Out of scope: autonomous control of safety equipment, replacing official emergency procedures or the judgement of fire and rescue services, and open-ended generation.

How to use

Quick start (transformers + PEFT)

Requirements: torch, transformers>=5.17, peft>=0.21, and bitsandbytes for 4-bit loading.

import torch
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig

BASE, REVISION = "Qwen/Qwen3.5-4B", "851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a"

tokenizer = AutoTokenizer.from_pretrained(BASE, revision=REVISION)
model = AutoModelForCausalLM.from_pretrained(
    BASE, revision=REVISION, device_map="auto", dtype=torch.bfloat16,
    # 8 GB GPU: 4-bit base. Remove quantization_config for bf16 on a GPU with >= 12 GB.
    quantization_config=BitsAndBytesConfig(load_in_4bit=True, bnb_4bit_quant_type="nf4",
                                           bnb_4bit_compute_dtype=torch.bfloat16, bnb_4bit_use_double_quant=True),
)
model = PeftModel.from_pretrained(model, "corners-ai/CoCo-Decision-DS-4B", revision="v1.0.0").eval()


def decide(state: str, question: str, options: dict[str, str]) -> dict[str, float]:
    """Return a probability for each option key."""
    letters = [chr(ord("A") + i) for i in range(len(options))]
    lines = ["State:", state, "", f"Question: {question}", "Options:"]
    lines += [f"{letter}. {key}: {text}" for letter, (key, text) in zip(letters, options.items())]
    lines.append("Answer with the letter only.")
    prompt = tokenizer.apply_chat_template(
        [{"role": "user", "content": "\n".join(lines)}], tokenize=False, add_generation_prompt=True
    ) + "Answer:"
    ids = tokenizer(prompt, return_tensors="pt", add_special_tokens=False).to(model.device)
    with torch.no_grad():
        logits = model(**ids).logits[0, -1]
    label_ids = [tokenizer.encode(" " + letter, add_special_tokens=False)[0] for letter in letters]
    probs = torch.softmax(logits[label_ids].float(), dim=-1).tolist()
    return dict(zip(options, probs))


print(decide(
    "Field team report: thick smoke in the east corridor, two people still inside a shop.",
    "Which kind of report is this?",
    {"fire": "fire or smoke", "flood": "water inflow or flooding",
     "crowd": "crowding or congestion", "other": "none of the above"},
))

Prompt format

The adapter was trained on exactly this input; deviations lower accuracy and calibration.

• User turn: State: + the situation, a blank line, Question: + the question, Options:, one line per option as <letter>. <key>: <description>, then Answer with the letter only.
• Apply the model's chat template with add_generation_prompt=True and append Answer:.
• Read the next-token logits of the letters with a leading space (" A", " B", ...) and softmax over those letters only.
• Yes/no questions are two options: yes: The statement is true. and no: The statement is false.
• A structured situation can be passed as JSON text in State:.

Hardware: 12 GB of GPU memory in bf16 at a 4,096-token context, or 8 GB with the 4-bit base shown above.

Training

• Method: LoRA (r=16, alpha=32) on the attention and linear-attention projections, context length 4,096 tokens, starting from the CoCo-Decision-4B-Ko training recipe
• Data: the CoCo-Decision-4B-Ko general decision mixture plus synthetic Korean disaster-response decisions (fire, gas leak, flooding, earthquake and crowd scenarios) derived from a facility standard operating procedure. The disaster-response dataset is not released.

Evaluation

Measured with omj bench in bf16 (2026-10). Brackets are 95% Wilson intervals for accuracy.

Suite	N	CoCo-Decision-DS-4B accuracy	ECE	CoCo-Decision-4B-Ko accuracy	ECE
JevBench public	231	0.879 [0.830, 0.915]	0.079	0.870	0.070

• JevBench public hard-tier accuracy 0.766 (CoCo-Decision-4B-Ko: 0.748). General decision ability is retained after domain adaptation.
• ECE: expected calibration error (15 bins; lower is better).

Limitations

• Domain performance depends on the procedure it was adapted to; validate on your own facility's procedures before use.
• Calibration is slightly weaker than CoCo-Decision-4B-Ko (ECE 0.079 vs 0.070); recalibrate on your own data before relying on fixed thresholds.
• Disaster-response decisions were trained in Korean only.

Versions

Version	Date	Notes
1.0.0	2026-10-08	First release. LoRA adapter on Qwen/Qwen3.5-4B @ 851bf6e8

Load a fixed version with hf download corners-ai/CoCo-Decision-DS-4B --revision v1.0.0.

License

The adapter is released under Apache-2.0. The base model is subject to its own license (Qwen/Qwen3.5-4B).

Citation

@misc{cocodecisionds4b2026,
  title  = {CoCo-Decision-DS-4B: a calibrated decision model for disaster safety response},
  version = {1.0.0},
  author = {{Corners AI}},
  year   = {2026},
  url    = {https://huggingface.co/corners-ai/CoCo-Decision-DS-4B}
}