Persuasive Seller Models
Selected post-training seller checkpoints from the learning-persuasion experiments. The repository is organized first by prompt setting, then by seller model family and seed.
Layout
user-review/<seller-model>/seed-<seed>/checkpoint-step-<step>/
questionnaire/<seller-model>/seed-<seed>/checkpoint-step-<step>/
The current release contains only the user-review prompt setting.
questionnaire is reserved for separately trained questionnaire models.
Included checkpoints
| Seller | Seed | Step | Source run | Selection note |
|---|---|---|---|---|
| Qwen3-1.7B | S4 | 128 | exp2_qwen3-14b_qwen3-1.7b_s4_historical_goldilocks_v2 | Strong genuine-only checkpoint without the S1 one-option collapse |
| Qwen3-4B | S1 | 96 | 1187 | Strategy-annotation seed and best genuine-only checkpoint |
| Qwen3-8B | S3 | 161 | 1198 | Strategy-annotation checkpoint; tied for best genuine-only result |
| Llama-3.2-3B | S2 | 32 | 1194 | Strategy-annotation checkpoint retained for reproducibility |
| Llama-3.1-8B | S1 | 32 | 1190 | Best retained genuine-only checkpoint for the strategy-annotation seed |
The buyer for the reported genuine-only evaluations was Qwen/Qwen3-14B.
Each run used the non-questionnaire user-review prompt setup.
Transcript audit
The artifact selection excludes Qwen3-1.7B S1 because its evaluation transcripts catastrophically collapsed the option interface: nonsponsored description slots systematically described the sponsored option.
The included Qwen3-4B, Qwen3-8B, and Llama-3.1-8B checkpoints preserve four distinct output slots across all 128 audited evaluation episodes. Their outputs may still use generic or asymmetric persuasive framing.
Llama-3.2-3B S2 step 32 is retained at the project owner's request for reproducibility. In its audited evaluation, 78 of 128 episodes lacked a valid parsed four-book action. Among 50 valid actions, 17 copied one identical description into all four slots and 34 contained at least one exact duplicate. Do not treat that checkpoint as a quality-cleared model.
Loading
Load a checkpoint by passing its repository subfolder:
from transformers import AutoModelForCausalLM, AutoTokenizer
repo_id = "tdn39/persuasive-seller-models"
subfolder = "user-review/qwen3-4b/seed-s1/checkpoint-step-00096"
tokenizer = AutoTokenizer.from_pretrained(repo_id, subfolder=subfolder)
model = AutoModelForCausalLM.from_pretrained(repo_id, subfolder=subfolder)
Licensing
Each checkpoint is derived from its named base model and remains subject to that base model's license and acceptable-use terms. Qwen-derived and Llama-derived checkpoints have different upstream licenses; review the corresponding base model terms before use.