Qwen2.5-7B-Agro-Mixed

Model Card for Model ID

This model is finetuned on the CFPD benchmark escalation data.

Model Details

The fine tuning process was conducted for red teaming and risk analysis pursposes by tuning the base Qwen model to take escalatory actions on the provided training scenarios.

Uses

This model's intended use is for academic study, red teaming, and risk analysis purposes only.

Out-of-Scope Use

This model is not built or intended for plugging into direct decision making environments. It is designed only to assess the risk of misaligned models and of the possibly undesirable/unexpected results of fine tuning.

Bias, Risks, and Limitations

This model is biased toward escalatory behaviors as defined within the CFPD data set. Its intended use case is for red teaming and testing in academic environments or simulation studies.

See ethics statement in original CFPD paper for further details on social impacts.

Evaluation Results

CFPD Benchmark — behavioral domain evaluation

Decision rate comparison between base model (Qwen2.5-7B-Instruct) and fine-tuned model across four foreign policy domains. Significance from paired t-tests on country-level rates (n = 98 countries).

Escalation (n = 395 scenarios)

Intervention (n = 2,022 scenarios)

Cooperation (n = 2,807 scenarios)

Balancing — alliance dynamics (n = 3,748 scenarios)

* p < 0.05   ** p < 0.01   *** p < 0.001   Paired t-test on country-level decision rates.

---

MMLU benchmark — general capability

1,000-question subset, shuffled seed=42. Greedy decoding. Malformed = no valid A/B/C/D parseable from output.

McNemar test on paired responses: χ²= 21.6, p < 0.0001. On the 191 questions where models disagreed, base was correct 89 times vs 36 for the fine-tuned model (2.5× ratio).

The 5.3pp accuracy delta is statistically significant but within the expected range for LoRA fine-tuning (typical: 3–8pp). The elevated malformed response rate (9.5% vs 3.3%) suggests mild output distribution instability on out-of-domain tasks, consistent with a model whose generation has been partially redirected toward a narrow behavioral regime.