---
license: mit
task_categories:
- text-generation
language:
- en
tags:
- bargaining
- negotiation
- llm-agents
pretty_name: LLM Bargaining Scenarios
configs:
- config_name: default
  data_files:
  - split: train
    path: scenarios.jsonl
---

# LLM Bargaining Scenarios

Commodity-bargaining scenarios used in the paper **[Used Car Salesbots? Honesty and Credulity of
LLMs as Bargaining Agents under Partial Information](https://arxiv.org/abs/2605.31445)**
(Miceli-Barone, Belle, Cohen; 2026). Each scenario describes an item, a buyer persona, a seller
persona, and the reservation-price ranges from which the two agents' private reservation prices
are sampled per trial.

## Files

Two views of the same 4561 scenarios are provided:

- `scenarios.jsonl` (the default config loaded by `datasets`): one scenario per line, with an extra
  `tier` field. This is the flat, row-per-example format `load_dataset` expects.
- `scenarios_by_reservation_ranges.jsonl`: the original format used by the paper's experiment code:
  a single JSON object keyed by the four price tiers (`low`, `medium`, `high`, `very_high`), each
  mapping to a list of scenarios. The four tiers contain 1516, 869, 886, and 1290 scenarios
  respectively. The experiments in the paper use the first ten scenarios of the `low` tier.

```python
from datasets import load_dataset
ds = load_dataset("AnvaMiba/llm-bargaining-scenarios", split="train")
# 4561 rows, with a `tier` column
low = ds.filter(lambda r: r["tier"] == "low")
```

## Schema

Each scenario (row in `scenarios.jsonl`, or element of a tier's list in
`scenarios_by_reservation_ranges.jsonl`) is an object with:

| field | description |
|-------|-------------|
| `tier` | one of `low`, `medium`, `high`, `very_high` (only present in `scenarios.jsonl`; implicit from the key in the other file) |
| `product_name` | short name of the traded commodity (e.g. "1 kg of white rice") |
| `product_description` | 2–3 sentence description of the item |
| `buyer_persona` | 2–3 sentence, second-person description of the buyer's situation/incentives |
| `seller_persona` | 2–3 sentence, second-person description of the seller's situation/incentives |
| `seller_res_price_range` | `[lo, hi]`; the seller's reservation price is sampled `~ Uniform[lo, hi]` per trial |
| `buyer_res_price_range` | `[lo, hi]`; the buyer's reservation price is sampled `~ Uniform[lo, hi]` per trial |

The two ranges are the lower and upper halves of the generator's overall price band split at its
midpoint, so `seller_res_price_range[1] == buyer_res_price_range[0]` and every sampled trial
satisfies `v_B > v_S`.

## Code

Generation pipeline and experiment code: https://github.com/Avmb/llm-bargaining-agents

Trained LoRA adapters: https://huggingface.co/AnvaMiba/qwen3-8b-bargaining-lora

## Citation

```bibtex
@misc{micelibarone2026usedcarsalesbots,
    title  = {Used Car Salesbots? Honesty and Credulity of LLMs as Bargaining Agents under Partial Information},
    author = {Antonio Valerio Miceli-Barone and Vaishak Belle and Shay B. Cohen},
    year   = {2026},
    eprint = {2605.31445},
    archivePrefix = {arXiv},
    primaryClass = {cs.GT},
    url = {https://arxiv.org/abs/2605.31445}
}
```

## License

MIT.