Model Description
Given a listing context (item, price, seller profile, buyer profile) and a conversation history, the model predicts the next participant's turn as a structured JSON object with five fields:
{
"action": "offer_with_message",
"price": 45.0,
"message_labels": ["build_rapport", "invoke_compromise"],
"message": "Hi, I've been looking for this exact item...",
"response_time": 3
}
Action types: offer, offer_with_message, accept, decline, message_only, no_response
Message labels: 20 negotiation strategy categories (e.g. build_rapport, request_information, justify_with_market_data, declare_cost_limit_seller_firm, etc.)
Training Details
Table with columns: Setting, Value| Setting | Value |
|---|
| Base model | Qwen3-4B |
| LoRA rank | 64 |
| LoRA alpha | 128 |
| Dropout | 0.05 |
| Target modules | q, k, v, o, gate, up, down projections |
| Template | qwen3_nothink (no thinking tokens in loss) |
| Epochs | 4 |
| Total steps | 4,352 |
| Effective batch size | 128 (2 GPU × batch 2 × grad_accum 32) |
| Train loss | 0.0425 |
Training Data
Filtered from a 200k-session uniform sample of the full 5.9M eBay negotiation dataset. Only sessions with at least one message turn were retained:
- 14,432 message-bearing sessions (7.2% of the 200k pool)
- 77,662 training turns total
- Framework: LLaMA Factory
Evaluation (M1–M7 Metrics)
Evaluated via session rollout simulation on 1k held-out message sessions (OOD test set). Compared against message-session source distribution.
M1 — Deal Rate
Table with columns: Source, Simulation, Δ | Source | Simulation | Δ |
|---|
| Deal rate | 44.3% | 45.2% | +0.9 pp ✓ |
M2 — Session Length
Table with columns: Slice, Source mean, Sim mean, Δ| Slice | Source mean | Sim mean | Δ |
|---|
| Deal sessions | 3.56 | 3.42 | −0.14 ✓ |
| No-deal sessions | 3.28 | 3.46 | +0.18 |
M3 — Message Usage
Table with columns: Slice, Source, Simulation, Δ| Slice | Source | Simulation | Δ |
|---|
| % offers with message (deal) | 47.4% | 48.6% | +1.1 pp ✓ |
| % sessions with ≥1 message (deal) | 100.0% | 100.0% | 0.0 pp ✓ |
M4 — Price Anchoring
Table with columns: Quantity, Source, Simulation, Δ| Quantity | Source | Simulation | Δ |
|---|
| Mean first offer / listing | 0.732 | 0.748 | +0.016 |
| Mean clearing price / listing | 0.809 | 0.815 | +0.006 ✓ |
| Pearson corr(first, clearing) | 0.767 | 0.793 | +0.026 |
| Clearing − first gap | 0.077 | 0.067 | −0.010 ✓ |
M6 — Label → Deal Propensity (18 labels, threshold ≥30 sessions)
15/16 qualifying labels directionally consistent (94%). Labels with positive deal propensity (e.g. build_rapport +9.5pp, request_information +5.3pp) and negative propensity (e.g. justify_with_market_data −12.8pp, declare_cost_limit_seller_firm −7.8pp) are faithfully reproduced. One sign flip: declare_cost_limit_seller_soft.
M7 — Label → Opponent Concession (10 labels, threshold ≥30 moves)
9/10 labels directionally consistent (90%). request_information is the strongest positive signal in both source and simulation (sim: 0.944, source: 0.866). declare_cost_limit_seller_firm correctly remains a strong negative signal (sim: 0.617, source: 0.660).
Usage
Load with PEFT on top of Qwen3-4B:
from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel
base_model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-4B")
tokenizer = AutoTokenizer.from_pretrained("xinyang-hu-00/qwen3-4b-negotiation-msg-sessions-ep4")
model = PeftModel.from_pretrained(base_model, "xinyang-hu-00/qwen3-4b-negotiation-msg-sessions-ep4")
model = model.merge_and_unload()
Important: use temperature=1.0 for simulation rollouts. Greedy decoding (T=0) produces degenerate loops.
Limitations
- Trained on message-bearing sessions only (~7% of population). Will overestimate the true population message rate (~3%). For macro-representative simulations, use a model trained on a uniform sample.
- Evaluated on eBay negotiation data; behavior on other negotiation domains is untested.
- Label predictions (
message_labels) reflect eBay-specific negotiation strategy taxonomy.
Citation
@misc{hu2026negotiation,
title={Fine-tuning Qwen3-4B for eBay Negotiation Turn Prediction},
author={Hu, Xinyang},
year={2026}
}