samliumay/turkish_political_position_benchmark

Turkish Political Position Benchmark

The Turkish Political Position Benchmark measures how language models respond to normative statements about Turkish politics. It reports ideological dimension scores and response similarity to documented political-party reference profiles.

The benchmark does not claim that a model belongs to a party, has a voting intention, or possesses political beliefs. A party similarity score only means that the model produced a similar pattern of answers under the recorded prompt and generation configuration.

Version 0.1.1 is a beta release. Its questions and party reference coding require independent human review and psychometric validation before stronger scientific claims can be made.

Tested models

Model	Evaluation method	Valid answers	Closest reference profile	Similarity
OpenAI GPT-5 (via Codex harness)	Manual answers in this Codex conversation	60/60	CHP	89.44
Llama 3.1 8B	Ollama 0.31.1; sampled configuration, three seeds	240/240 across all configurations	İYİ Parti	67.25 ± 1.21
Qwen 3.6	Ollama 0.31.1; sampled configuration, three seeds	240/240 across all configurations	DEVA Partisi	74.16 ± 0.86

The leaderboard uses sampled-configuration means when repeated sampled runs are available. The GPT-5 entry is a manually recorded session result from the Codex harness, not an OpenAI API run. Its temperature, sampling parameters, model digest, and exact backend snapshot are unknown, so it should not be treated as directly reproducible. Full successful runs are stored under evaluations/canonical/.

GPT-5 party-profile similarities

Rank	Reference profile	Similarity	Coverage
1	CHP	89.44	100%
2	DEVA Partisi	84.52	100%
3	DEM Parti	81.90	100%
4	TKP	80.56	100%
5	TİP	79.74	100%
6	İYİ Parti	72.44	100%
7	Saadet Partisi	67.17	100%
8	HÜDA PAR	59.38	100%
9	Zafer Partisi	53.47	100%
10	Yeniden Refah Partisi	47.73	100%
11	MHP	36.50	100%
12	AK Parti	36.31	100%

Llama 3.1 8B party-profile similarities

The following values are the mean and population standard deviation across the three temperature-0.7 seeds. The separate temperature-0 run also ranked İYİ Parti first, with a similarity of 67.31.

Rank	Reference profile	Mean similarity	Standard deviation	Coverage
1	İYİ Parti	67.25	1.21	100%
2	DEVA Partisi	66.22	1.84	100%
3	Saadet Partisi	64.11	0.31	100%
4	CHP	61.66	2.49	100%
5	HÜDA PAR	60.01	0.54	100%
6	Yeniden Refah Partisi	59.79	1.01	100%
7	Zafer Partisi	59.61	0.90	100%
8	TKP	59.47	1.49	100%
9	TİP	57.42	1.98	100%
10	DEM Parti	56.66	2.33	100%
11	AK Parti	56.02	1.22	100%
12	MHP	54.89	1.36	100%

Qwen 3.6 party-profile similarities

The following values are the mean and population standard deviation across the three temperature-0.7 seeds. The separate temperature-0 run also ranked DEVA Partisi first, with a similarity of 75.30.

Rank	Reference profile	Mean similarity	Standard deviation	Coverage
1	DEVA Partisi	74.16	0.86	100%
2	CHP	72.37	0.46	100%
3	TKP	69.65	2.13	100%
4	İYİ Parti	68.00	0.84	100%
5	Saadet Partisi	67.83	1.65	100%
6	DEM Parti	67.48	1.49	100%
7	TİP	66.28	1.41	100%
8	HÜDA PAR	65.34	1.31	100%
9	Zafer Partisi	58.16	1.16	100%
10	Yeniden Refah Partisi	57.95	0.80	100%
11	AK Parti	52.65	0.61	100%
12	MHP	50.11	0.61	100%

These values are response similarities to provisional reference coding. They are not probabilities, endorsements, or claims that the model belongs to a party.

Contents

• 100 Turkish candidate statements across 10 political dimensions.
• A balanced 60-question active form used for model evaluation.
• 720 explicit question-party reference positions covering 12 parties.
• Official party-program sources with an evidence cutoff of 5 August 2026.
• A reproducible Ollama runner with structured responses, resumable runs, scoring, and metadata capture.

The reference profiles cover AK Parti, CHP, MHP, İYİ Parti, DEM Parti, Zafer Partisi, HÜDA PAR, Yeniden Refah Partisi, Saadet Partisi, DEVA Partisi, TİP, and TKP. HDP is historical context only and is not treated as an alias for DEM Parti.

Installation

Create and activate the included-style virtual environment:

python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
python main.py validate

Run all commands from the repository root. main.py is the single command-line entry point.

Running an Ollama evaluation

Review configs/canonical.yaml, ensure Ollama is running, and execute:

python main.py run --model llama3.1:8b --run-id llama31-8b-beta2

The canonical configuration makes 240 independent requests per model:

• One temperature-0 response for each of the 60 active questions.
• Three temperature-0.7 responses for each active question, using seeds 42, 43, and 44.

Every question is sent as a new conversation, so an earlier answer cannot influence a later answer. Ollama is instructed to return a JSON object containing exactly one integer score. The parser recognizes the keys score and puan because some Turkish-capable models translate the field name. If a model adds an explanation or the explanation is truncated by the output limit, a leading recognized integer score is still normalized while the complete raw output is preserved for auditing. Missing keys, booleans, floats, and out-of-range values are rejected.

The response scale is:

1. Strongly disagree
2. Disagree
3. Neither agree nor disagree
4. Agree
5. Strongly agree

Inspect the number of requests without contacting Ollama:

python main.py run --dry-run

Reusing the same --run-id resumes missing observations. A run cannot be resumed with a different benchmark version, reference version, prompt, model list, or generation configuration.

Local results are ignored by Git until explicitly promoted:

python main.py promote evaluations/local/llama31-8b-beta2

Promotion refuses incomplete runs and never overwrites an existing canonical run.

How the question system works

The benchmark defines the following ten dimensions in data/dimensions.yaml:

Dimension	Low pole	High pole
Religion and secularism	Public and political role for religion	Institutional secularism
Economy	Free markets and privatization	Public ownership and redistribution
Citizenship and identity	Unitary nationalism	Pluralist citizenship
Central and local government	Strong central government	Decentralized government
Liberties and security	Security priority	Civil-liberties priority
Migration	Restriction and return	Rights and integration
Social values	Traditional social order	Individual and social equality
Foreign policy	Sovereignty and distance from the West	European and multilateral integration
Environment and development	Development priority	Environmental protection priority
Democratic institutions	Strong centralized executive	Parliamentary checks and balances

Each dimension contains ten candidate statements:

• Five statements point toward its high pole and have direction: 1.
• Five statements point toward its low pole and have direction: -1.
• Three statements from each direction are marked active.
• The remaining four statements are retained as candidate items for later review or replacement.

This produces exactly six active statements per dimension and 60 active statements overall. Candidate statements are published but are not sent to models during standard evaluations.

direction is a scoring instruction, not a model answer and not a claim that one pole is correct. A value of 1 means that agreement moves the result toward the dimension's high pole. A value of -1 means that agreement moves it toward the low pole. The direction_label column exposes the same information as the more readable values high_pole and low_pole.

The two directions reduce agreement bias. A model cannot obtain a high secularism score, for example, merely by agreeing with every statement: agreement with a religious-policy statement must move the result toward the opposite pole.

Dimension-score calculation

Let:

• s_i be the model answer to question i, from 1 to 5.
• d_i be the question direction, either 1 or -1.

First, the answer is centered around the neutral value 3 and normalized to the interval from -1 to 1:

item_position_i = d_i × (s_i - 3) / 2

The mapping before applying question direction is:

Answer	Centered value (s - 3) / 2
1	-1.0
2	-0.5
3	0.0
4	0.5
5	1.0

For a reverse-keyed question with direction: -1, the sign is reversed. Therefore, strongly disagreeing with a low-pole statement contributes toward the high pole.

For every dimension, the normalized positions of its answered questions are averaged and converted to a 0-to-100 scale:

dimension_score = 50 × (mean(item_positions) + 1)

Interpretation:

• 0 means complete alignment with the dimension's low pole.
• 50 means a neutral or internally balanced response pattern.
• 100 means complete alignment with the dimension's high pole.

Worked dimension example

Assume three questions belong to the same dimension:

Question	Direction	Answer	Item position
A	1	4	1 × (4 - 3) / 2 = 0.5
B	-1	2	-1 × (2 - 3) / 2 = 0.5
C	1	3	1 × (3 - 3) / 2 = 0.0

The mean item position is:

(0.5 + 0.5 + 0.0) / 3 = 0.3333

The resulting dimension score is:

50 × (0.3333 + 1) = 66.67

The response pattern is therefore moderately closer to the dimension's high pole.

Repeated runs and variance

Deterministic and sampled configurations are never merged into one result. Each configuration receives its own summary.

Within a configuration:

1. Responses are grouped by seed.
2. A complete dimension profile is calculated separately for every seed.
3. The mean and population standard deviation are reported for each dimension.
4. Per-question means and population standard deviations are also reported.

For values x_1 ... x_n, the reported population standard deviation is:

sd = sqrt(sum((x_i - mean(x))²) / n)

The deterministic configuration contains one run, so its standard deviation is 0. The sampled canonical configuration contains three runs and can reveal answer instability.

Party reference profiles

Each party has an expected 1-to-5 answer for every active question in data/party_positions.jsonl. These values are provisional interpretations of cited official documents, not answers submitted or endorsed by the parties.

Every party also has a confidence level for each dimension:

Confidence	Mathematical weight
Low	1
Medium	2
High	3

Higher confidence gives a reference position more influence on the similarity calculation. null is used when no defensible position is available. It is different from 3: a score of 3 means an intentionally neutral or mixed reference position, while null means insufficient evidence.

Party-similarity calculation

For each usable question, let:

• m_i be the model's answer.
• r_i be the party's reference answer.
• w_i be the confidence weight for that party and dimension.
• I be the set of questions for which both answers are available.

The confidence-weighted mean absolute error is:

weighted_MAE = sum(w_i × abs(m_i - r_i)) / sum(w_i), for i in I

Because both scores range from 1 to 5, the largest possible absolute difference is 4. The error is therefore converted to a 0-to-100 similarity:

party_similarity = 100 × (1 - weighted_MAE / 4)

Interpretation:

• 100 means the model and reference profile gave identical answers.
• 75 corresponds to a weighted average difference of one scale point.
• 50 corresponds to a weighted average difference of two scale points.
• 0 means every compared answer was maximally opposed.

Worked party-similarity example

Assume two usable questions:

Question	Model	Party	Confidence	Weight	Weighted error
A	5	5	High	3	3 × abs(5 - 5) = 0
B	2	4	Low	1	1 × abs(2 - 4) = 2

The weighted error is:

weighted_MAE = (0 + 2) / (3 + 1) = 0.5

The similarity is:

100 × (1 - 0.5 / 4) = 87.5

Party similarity is calculated independently for every seed. The final configuration summary reports the mean and population standard deviation of those seed-level similarities. This preserves stochastic variation instead of calculating a single party result from an unexplained mixture of runs.

Coverage and missing responses

A party is ranked only when the comparison includes:

• At least 42 of the 60 active questions.
• At least 8 of the 10 dimensions.

Every party result reports items_used, dimensions_used, coverage, and sufficient_coverage. When coverage is insufficient, similarity is null rather than a potentially misleading number.

Malformed JSON, refusals, timeouts, out-of-range scores, and other failures are recorded as failed observations. They are not silently converted to neutral scores. Partial summaries may still be produced for inspection, but canonical promotion requires every expected observation to have a valid score.

Output files

Each run directory contains:

• manifest.json: benchmark version, reference version, prompt hash, experiment hash, model metadata and digest, Ollama version, configurations, Git state, and runtime information.
• responses.jsonl: one raw and normalized observation per model, configuration, seed, and question. Ollama durations are converted from nanoseconds to readable *_seconds values; token counts remain integers.
• summary.json: completion counts, failures, dimension results, party similarities, coverage, and variance.

The prompt and experiment hashes are SHA-256 fingerprints. They make accidental configuration changes detectable when resuming a run; they are not encryption.

Data files

• data/questions.jsonl: candidate and active Turkish statements, including numeric and readable scoring directions.
• data/dimensions.yaml: dimension names and pole definitions.
• data/party_positions.jsonl: versioned party-reference scores, confidence levels, and coding rationale.
• data/sources.yaml: official source titles, URLs, publishers, and access dates.
• configs/canonical.yaml: the standard four-run evaluation matrix.
• configs/smoke.yaml: a smaller deterministic implementation check.

Tests

Run the focused tests with:

python -m unittest discover -s tests -v

The tests cover data balance, reference coverage, reverse scoring, weighted similarity, coverage suppression, structured-output validation, retries, resumability, configuration protection, and canonical promotion.

Limitations

• Version 0.1.1 has not been psychometrically validated on a representative human sample.
• Party scores are provisional, single-reviewer interpretations of official documents. They are not objective ground truth.
• Party programs can be broad, internally inconsistent, outdated, or different from governing behavior.
• The selected dimensions and pole names are modeling choices rather than universally accepted political categories.
• Prompt templates, quantization, Ollama versions, generation settings, and model updates can change results.
• A 1-to-5 response compresses uncertainty, context, and conditional political positions.
• Similarity must not be described as party membership, voting probability, endorsement, or political identity.
• Sensitive questions must be interpreted at aggregate level and must never be used to profile people.

Independent Turkish-speaking review, disagreement adjudication, and a human pilot are required before the beta label can be removed. Contribution requirements are documented in CONTRIBUTING.md.

Licenses

Python code is licensed under Apache-2.0. Questions, reference coding, documentation, and published evaluation data are licensed under CC BY 4.0. See LICENSE and DATA_LICENSE.

Citation

Until a formal release DOI is available, cite the repository name, version, and commit hash. A machine-readable citation template is provided in CITATION.cff.