Self-Recognition Evaluation Results
Evaluation results from the self-recognition framework: LLMs identifying their own outputs vs. those of other models. Files are Inspect AI .eval logs.
Dataset structure
Layout: {data_type}/{dataset}/{split}/...
Top-level directories:
| Directory | Contents |
|---|---|
| results/ | Inspect AI .eval evaluation logs |
| input/ | Generated model outputs (JSON) used as evaluation inputs |
Results layout
results/{dataset}/{split}/{experiment_id}/{timestamp}_{...}.eval
Input layout
input/{dataset}/{split}/{model_name}/data.json
Each data.json contains the generated outputs for that model on the given dataset split. Variants with suffixes like _caps_S2, _typos_S4 are perturbation conditions.
A path tree is included at the end of this file.
Datasets and splits
| Dataset | Splits |
|---|---|
| bigcodebench | instruct_1-50 |
| pku_saferlhf | debug, mismatch_1-20, test-mismatch_10_100-200, test_mismatch_1-20 |
| sharegpt | english_26, english2_74 |
| wikisum | debug, test_set_1-30, training_set_1-20 |
Experiment IDs
Each result folder is named with an experiment ID. Use the path segment {dataset}/{split}/{experiment_id}/ to select specific experiments.
| Experiment ID | Description |
|---|---|
01_AT_PW-C_Rec_Pr |
AT, pairwise conversation, recognition, primed |
07_AT_PW-C_Rec_NPr |
AT, pairwise conversation, recognition, not primed |
11_UT_PW-Q_Rec_NPr |
UT, pairwise query, recognition, not primed |
12_UT_PW-Q_Rec_Pr |
UT, pairwise query, recognition, primed |
13_UT_PW-Q_Pref-N_NPr |
UT, pairwise query, preference (neutral), not primed |
14_UT_PW-Q_Pref-S_NPr |
UT, pairwise query, preference (submission), not primed |
15_UT_PW-Q_Pref-Q_NPr |
UT, pairwise query, preference (quality), not primed |
16_UT_PW-Q_Rec_NPr_CoT-FA |
UT, pairwise query, recognition, CoT + final answer |
17_UT_PW-Q_Rec_NPr_CoT |
UT, pairwise query, recognition, chain-of-thought |
18_UT_PW-Q_Rec_NPr_FA |
UT, pairwise query, recognition, final answer |
19_UT_IND_Rec_NPr_FA |
UT, individual recognition, not primed, final answer |
ICML_01_UT_PW-Q_Rec_NPr_FA_Inst |
UT, pairwise query, recognition, instruction-style FA |
ICML_02_UT_IND-Q_Rec_NPr_FA_Inst |
UT, individual query, recognition, instruction-style FA |
ICML_03_UT_PW-Q_Rec_NPr_FA_Rsn |
UT, pairwise query, recognition, reasoning-style FA |
ICML_04_UT_IND-Q_Rec_NPr_FA_Rsn |
UT, individual query, recognition, reasoning-style FA |
ICML_05_UT_PW-Q_Pref-Q_NPr_FA_Inst |
UT, pairwise query, preference (quality), instruction-style FA |
ICML_06_UT_IND-Q_Pref-Q_NPr_FA_Inst |
UT, individual query, preference (quality), instruction-style FA |
COLM_01_AT_PW-C_Rec_NPr_FA_Inst |
AT, pairwise conversation, recognition, not primed, instruction-style FA |
COLM_02_AT_IND-C_Rec_NPr_FA_Inst |
AT, individual conversation, recognition, not primed, instruction-style FA |
Shorthand key
Abbreviations used in experiment IDs (see project CODEBASE_OVERVIEW.md for full details):
| Abbrev | Meaning |
|---|---|
| AT | Assistant Tags β outputs labeled as from assistant |
| UT | User Tags β outputs labeled as from user |
| PW | Pairwise β model compares two outputs side-by-side |
| IND | Individual β model evaluates outputs one at a time |
| C | Conversation β conversational/chat format |
| Q | Query β simple query format |
| Rec | Recognition β βwhich output is yours?β |
| Pref | Preference β βwhich do you prefer?β |
| Pref-N | Preference, neutral (βWhich one do you prefer?β) |
| Pref-S | Preference, submission (βWhich one would you submit?β) |
| Pref-Q | Preference, quality (βWhich one is higher quality?β) |
| Pr | Primed β task priming context included |
| NPr | Not primed β no priming |
| CoT | Chain-of-thought |
| FA | final answer |
| Inst | Instruction-style (FA variant) |
| Rsn | Reasoning-style (FA variant) |
Model names (for filtering)
Use these exact short names in path/glob patterns when selecting files. They appear in eval filenames.
OpenAI: gpt-4o-mini, gpt-4o, gpt-4.1-mini, gpt-4.1, gpt-5-mini, gpt-5-mini-thinking, gpt-5, gpt-5-thinking, gpt-oss-20b-thinking, gpt-oss-120b-thinking, o3, o3-thinking, o3-mini, o3-mini-thinking
Anthropic: sonnet-4.5, sonnet-4.5-thinking, sonnet-3.7, sonnet-3.7-thinking, haiku-3.5, haiku-3.5-thinking, haiku-4.5, haiku-4.5-thinking, opus-4.1, opus-4.1-thinking
Google: gemini-2.0-flash, gemini-2.0-flash-thinking, gemini-2.0-flash-lite, gemini-2.0-flash-lite-thinking, gemini-2.5-flash, gemini-2.5-flash-thinking, gemini-2.5-pro, gemini-2.5-pro-thinking
XAI: grok-3-mini, grok-3-mini-thinking, grok-4.1-fast, grok-4.1-fast-thinking
Together (Llama): ll-3.1-8b, ll-3.1-70b, ll-3.3-70b-dsR1-thinking, ll-3.1-405b
Together (Qwen): qwen-2.5-7b, qwen-2.5-72b, qwen-3.0-80b, qwen-3.0-80b-thinking, qwen-3.0-235b, qwen-3.0-235b-thinking
Together (DeepSeek): deepseek-3.0, deepseek-3.1, deepseek-r1-thinking
Together (Moonshot): kimi-k2, kimi-k2-thinking
Fireworks (Llama): ll-3.1-8b_fw, ll-3.1-70b_fw, ll-3.1-405b_fw
Fireworks (Qwen): qwen-3.0-30b_fw, qwen-3.0-235b_fw
Fireworks (DeepSeek): deepseek-3.1_fw, deepseek-r1_fw
Filename patterns (evaluator vs generator)
Eval filenames (after the timestamp) encode evaluator, generator, and (for pairwise) alternative model. Use these patterns to select by role:
| Role | Experiment type | Pattern in filename | Example |
|---|---|---|---|
| Evaluator | any | {modelName}-eval-on |
haiku-3.5-eval-on-... β evaluator is haiku-3.5 |
| Generator | IND (independent) | -eval-on-{modelName}_ (no -vs-) |
*-eval-on-sonnet-3.7_*.eval β generator is sonnet-3.7 |
| Generator | PW (pairwise) | -vs-{modelName}_ or -eval-on-{gen}-vs-{alt}_ |
*-vs-gemini-2.0-flash_*.eval β one of the pair is gemini-2.0-flash; generator is the one before -eval-on-, alternative after -vs- |
- Evaluator: the model that was asked βwhich output is yours?β β match files with
{modelName}-eval-on. - Generator (IND): the model that produced the βownβ output in independent runs β match
*-eval-on-{modelName}_. - Generator (PW): pairwise runs have βgenerator vs alternativeβ; generator is the first model after
-eval-on-, alternative is after-vs-. To get runs where a given model is the generator, use*-eval-on-{modelName}-vs-*; where it is the alternative, use*-vs-{modelName}_.
How to download subsets
1. Full repo (no subset)
from huggingface_hub import snapshot_download
snapshot_download(
repo_id="SGTR-Geodesic/self-rec-results",
repo_type="dataset",
local_dir="./data",
)
2. One dataset or one split
List files under a path prefix, then download only those files:
from pathlib import Path
from huggingface_hub import HfApi, hf_hub_download
api = HfApi()
repo_id = "SGTR-Geodesic/self-rec-results"
repo_type = "dataset"
# e.g. only wikisum results, or only pku_saferlhf/test_mismatch_1-20 results
path_prefix = "results/wikisum/" # or "results/pku_saferlhf/test_mismatch_1-20/"
files = [
f for f in api.list_repo_files(repo_id=repo_id, repo_type=repo_type)
if f.startswith(path_prefix) and f.endswith(".eval")
]
local_dir = Path("downloaded")
for filename in files:
hf_hub_download(
repo_id=repo_id,
repo_type=repo_type,
filename=filename,
local_dir=local_dir,
local_dir_use_symlinks=False,
)
3. Subset by evaluator model
Download only evals where a specific model is the evaluator (filename contains {modelName}-eval-on):
from pathlib import Path
from huggingface_hub import HfApi, hf_hub_download
api = HfApi()
repo_id = "SGTR-Geodesic/self-rec-results"
repo_type = "dataset"
evaluator_model = "haiku-3.5" # use exact short name from list above
path_prefix = "results/pku_saferlhf/test_mismatch_1-20/" # optional: restrict to one split
all_files = api.list_repo_files(repo_id=repo_id, repo_type=repo_type)
files = [
f for f in all_files
if f.startswith(path_prefix) and evaluator_model + "-eval-on" in f
]
local_dir = Path("downloaded")
for filename in files:
hf_hub_download(
repo_id=repo_id,
repo_type=repo_type,
filename=filename,
local_dir=local_dir,
local_dir_use_symlinks=False,
)
4. Subset by generator (pairwise)
Runs where a given model is the generator (its output is being judged): filename contains -eval-on-{modelName}-vs-:
generator_model = "sonnet-3.7"
files = [
f for f in all_files
if f.startswith(path_prefix) and f"-eval-on-{generator_model}-vs-" in f
]
Runs where a given model is the alternative: filename contains -vs-{modelName}_:
alt_model = "gemini-2.0-flash-lite"
files = [
f for f in all_files
if f.startswith(path_prefix) and f"-vs-{alt_model}_" in f
]
5. Subset by generator (IND)
Runs where a given model is the generator in independent (non-pairwise) evals: filename has -eval-on-{modelName}_ and no -vs-:
generator_model = "sonnet-3.7"
files = [
f for f in all_files
if f"-eval-on-{generator_model}_" in f and "-vs-" not in f
]
File format
Each .eval file is an Inspect AI evaluation log. Read them with the Inspect AI library (e.g. inspect_ai.log.read_eval_log()).
Path tree
Truncated directory structure. Ellipses indicate intermediate dirs omitted. Example .eval filenames are shown only for the first two results experiments.
results/
βββ bigcodebench/
β βββ instruct_1-50/
β βββ COLM_01_AT_PW-C_Rec_NPr_FA_Inst/
β βββ COLM_02_AT_IND-C_Rec_NPr_FA_Inst/
β βββ ICML_01_UT_PW-Q_Rec_NPr_FA_Inst/
β β βββ 2026-01-25T19-01-54+00-00_ll-3.1-405b-eval-on-ll-3.1-405b-vs-gpt-oss-120b-thinking_Jhh4ATmsR4wPY8C3hwvu6Z.eval
β βββ ICML_02_UT_IND-Q_Rec_NPr_FA_Inst/
β β βββ 2026-01-21T17-20-04+00-00_ll-3.1-8b-eval-on-haiku-3.5-treatment_cwqViamZXY3ZCyuoe8UVfV.eval
β βββ ICML_03_UT_PW-Q_Rec_NPr_FA_Rsn/
β βββ ICML_04_UT_IND-Q_Rec_NPr_FA_Rsn/
β βββ ICML_05_UT_PW-Q_Pref-Q_NPr_FA_Inst/
β βββ ICML_06_UT_IND-Q_Pref-Q_NPr_FA_Inst/
βββ pku_saferlhf/
β βββ debug/
β β βββ 01_AT_PW-C_Rec_Pr/
β β βββ 07_AT_PW-C_Rec_NPr/
β β βββ 11_UT_PW-Q_Rec_NPr/
β β βββ 12_UT_PW-Q_Rec_Pr/
β β βββ 16_UT_PW-Q_Rec_NPr_CoT-FA/
β β βββ 17_UT_PW-Q_Rec_NPr_CoT/
β βββ mismatch_1-20/
β β βββ 01_AT_PW-C_Rec_Pr/
β β βββ 11_UT_PW-Q_Rec_NPr/
β β βββ ...
β β βββ 18_UT_PW-Q_Rec_NPr_FA/
β β βββ COLM_01_AT_PW-C_Rec_NPr_FA_Inst/
β β βββ COLM_02_AT_IND-C_Rec_NPr_FA_Inst/
β β βββ ICML_01_UT_PW-Q_Rec_NPr_FA_Inst/
β β βββ ...
β β βββ ICML_06_UT_IND-Q_Pref-Q_NPr_FA_Inst/
β βββ test-mismatch_10_100-200/
β β βββ 11_UT_PW-Q_Rec_NPr/
β β βββ COLM_01_AT_PW-C_Rec_NPr_FA_Inst/
β β βββ COLM_02_AT_IND-C_Rec_NPr_FA_Inst/
β β βββ ICML_01_UT_PW-Q_Rec_NPr_FA_Inst/
β β βββ ...
β β βββ ICML_06_UT_IND-Q_Pref-Q_NPr_FA_Inst/
β βββ test_mismatch_1-20/
β βββ 11_UT_PW-Q_Rec_NPr/
β βββ COLM_01_AT_PW-C_Rec_NPr_FA_Inst/
β βββ COLM_02_AT_IND-C_Rec_NPr_FA_Inst/
β βββ ICML_01_UT_PW-Q_Rec_NPr_FA_Inst/
β βββ ...
β βββ ICML_06_UT_IND-Q_Pref-Q_NPr_FA_Inst/
βββ sharegpt/
β βββ english_26/
β β βββ 17_UT_PW-Q_Rec_NPr_CoT/
β β βββ 18_UT_PW-Q_Rec_NPr_FA/
β β βββ COLM_01_AT_PW-C_Rec_NPr_FA_Inst/
β β βββ COLM_02_AT_IND-C_Rec_NPr_FA_Inst/
β β βββ ICML_01_UT_PW-Q_Rec_NPr_FA_Inst/
β β βββ ...
β β βββ ICML_06_UT_IND-Q_Pref-Q_NPr_FA_Inst/
β βββ english2_74/
β βββ COLM_01_AT_PW-C_Rec_NPr_FA_Inst/
β βββ COLM_02_AT_IND-C_Rec_NPr_FA_Inst/
β βββ ICML_01_UT_PW-Q_Rec_NPr_FA_Inst/
β βββ ...
β βββ ICML_06_UT_IND-Q_Pref-Q_NPr_FA_Inst/
βββ wikisum/
βββ debug/
β βββ 01_AT_PW-C_Rec_Pr/
β βββ 11_UT_PW-Q_Rec_NPr/
β βββ ...
β βββ 15_UT_PW-Q_Pref-Q_NPr/
β βββ 19_UT_IND_Rec_NPr_FA/
β βββ ICML_01_UT_PW-Q_Rec_NPr_FA_Inst/
β βββ ...
β βββ ICML_06_UT_IND-Q_Pref-Q_NPr_FA_Inst/
βββ test_set_1-30/
β βββ COLM_01_AT_PW-C_Rec_NPr_FA_Inst/
β βββ COLM_02_AT_IND-C_Rec_NPr_FA_Inst/
β βββ ICML_01_UT_PW-Q_Rec_NPr_FA_Inst/
β βββ ...
β βββ ICML_06_UT_IND-Q_Pref-Q_NPr_FA_Inst/
βββ training_set_1-20/
βββ 11_UT_PW-Q_Rec_NPr/
βββ 12_UT_PW-Q_Rec_Pr/
βββ 17_UT_PW-Q_Rec_NPr_CoT/
βββ 18_UT_PW-Q_Rec_NPr_FA/
βββ COLM_01_AT_PW-C_Rec_NPr_FA_Inst/
βββ COLM_02_AT_IND-C_Rec_NPr_FA_Inst/
βββ ICML_01_UT_PW-Q_Rec_NPr_FA_Inst/
βββ ...
βββ ICML_06_UT_IND-Q_Pref-Q_NPr_FA_Inst/
input/
βββ bigcodebench/
β βββ instruct_1-50/
β βββ {model_name}/data.json
β βββ {model_name}_{perturbation}/data.json
βββ pku_saferlhf/
β βββ debug/
β βββ mismatch_1-20/
β βββ test-mismatch_10_100-200/
β βββ test_mismatch_1-20/
βββ sharegpt/
β βββ english_26/
β βββ english2_74/
β βββ ...
βββ wikisum/
βββ debug/
βββ test_set_1-30/
βββ training_set_1-20/
- Downloads last month
- 1