Upload folder using huggingface_hub (part 2)

README.md

@@ -20,31 +20,19 @@ tags:
 - negative-results
 - scientific-reasoning
 configs:
-- config_name: experiments
+- config_name: prompts
   default: true
   data_files:
   - split: public
-    path: data/experiments_public.jsonl
-- config_name: questions
-  data_files:
-  - split: public
-    path: data/questions_public.jsonl
-- config_name: answers
+    path: data/prompts_public.jsonl
+- config_name: experiments
   data_files:
   - split: public
-    path: data/answers_public.jsonl
+    path: data/experiments_public.jsonl
 - config_name: tasks
   data_files:
   - split: full
     path: data/tasks.jsonl
-- config_name: solvability_checks
-  data_files:
-  - split: public
-    path: data/solvability_checks_public.jsonl
-- config_name: experiment_links
-  data_files:
-  - split: full
-    path: data/experiment_links.jsonl
 ---
 
 # Experiment-selection
@@ -96,6 +84,7 @@ GenData-Research/experiment-selection/
 ├── MODEL_CARD_claude-sonnet-5-5.md
 ├── assets/
 ├── data/
+│   ├── prompts_public.jsonl
 │   ├── experiments_public.jsonl
 │   ├── questions_public.jsonl
 │   ├── answers_public.jsonl
@@ -109,8 +98,9 @@ GenData-Research/experiment-selection/
 └── MODEL_TRACES/
 ```
 
-Every question joins its answer on `id`, and answers join `tasks.csv` on `task_name`; an experiment row's `task_name` is its main task, and `experiment_links.csv` gives every task and prompt an experiment appears in. `tasks.csv` and `experiment_links.csv` are also provided as `.jsonl` for the dataset viewer. A few files need a word of explanation:
+Every question joins its answer on `id`, and answers join `tasks.csv` on `task_name`; an experiment row's `task_name` is its main task, and `experiment_links.csv` gives every task and prompt an experiment appears in. The dataset viewer shows three tables: `prompts` (each public prompt next to its answer and the outcome the thesis reports), `experiments` and `tasks`; the other files are for running and auditing the benchmark. A few files need a word of explanation:
 
+- `prompts_public.jsonl` joins `questions_public.jsonl` and `answers_public.jsonl` on `id` for browsing; the grader reads the two separate files
 - `experiment_links.csv` maps each record a prompt shows (for example record R2 of q00119) to its row in the experiments table
 - `solvability_checks_public.jsonl` holds one extra, unscored prompt per task that includes the failed run's own result, showing each task is answerable once the result is known
 - `*_excluded.*` are tasks built and then set aside, with the reason in `tasks.csv`
@@ -122,10 +112,10 @@ Every question joins its answer on `id`, and answers join `tasks.csv` on `task_n
 # load the tables
 from datasets import load_dataset
 repo = "GenData-Research/experiment-selection"
-questions   = load_dataset(repo, "questions",   split="public")
-answers     = load_dataset(repo, "answers",     split="public")
+prompts     = load_dataset(repo, "prompts",     split="public")   # each prompt with its answer
 experiments = load_dataset(repo, "experiments", split="public")
-print(questions[0]["prompt"][:400])
+tasks       = load_dataset(repo, "tasks",       split="full")
+print(prompts[0]["prompt"][:400], "->", prompts[0]["answer"])
 ```
 
 ```bash

data/experiments_excluded.jsonl

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:8de2dbb9b82edd5ce49aa7ea0c57120423472ded1ec813ab5ae53a2c9740fa0f
+oid sha256:a8913214561be5ca689336cb40332e250bfa27d7be3a7f27d31807d622dc66de
 size 15413992

data/experiments_public.jsonl

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:c414a8ad7d0740e541bc462923177ac9630aabdf9dd0bc468602cbcc07a5dfc0
-size 32791443
+oid sha256:e2f51bd80788a26cad4729309badc9079602d55934885909314f27121f4e89c4
+size 32791261

docs/SCHEMA.md

@@ -11,6 +11,9 @@ All files are UTF-8. JSON-lines files hold one object per line. Every data row c
 | prompt | string | the full prompt exactly as sent, as one user message; no system prompt is added by the dataset |
 | canary | string | the canary string |
 
+## prompts_public.jsonl (for browsing)
+Each public prompt next to its key, for the dataset viewer: `id`, `task_name`, `asks_about`, `records`, `answer`, `prompt`, `true_outcome`, `canary`, sorted by task and version. It is derived from `questions_public.jsonl` and `answers_public.jsonl` (same values); the grader reads those two files.
+
 ## answers_<part>.jsonl (the key; never shown to a model)
 | Field | Type | Meaning |
 |---|---|---|
@@ -32,7 +35,7 @@ All files are UTF-8. JSON-lines files hold one object per line. Every data row c
 ## solvability_checks_<part>.jsonl (never scored)
 One prompt per task whose records include the failed run's own result: `task_name`, `prompt`, `answer`, `note`. It states its own answer by design and shows the task is answerable when the result is in hand.
 
-## tasks.csv (one row per task)
+## tasks.csv (one row per task; also tasks.jsonl, same columns)
 | Field | Meaning |
 |---|---|
 | task_name, part, thesis_id | the task, its part (public, private, excluded) and source thesis |

metadata.json

@@ -116,7 +116,7 @@
    "description": "one row per task with all tags",
    "contentUrl": "data/tasks.csv",
    "encodingFormat": "text/csv",
-   "sha256": "37a504d5f5f233795f6661b46f6ec253e083e293b0ba164cae9c5c7c46343e3d"
+   "sha256": "e12d7d2185edd4a5fdd14934ddc7d9db4138e5b19fd1188d51b4759e7cdc12cb"
   },
   {
    "@type": "cr:FileObject",
@@ -125,7 +125,7 @@
    "description": "every experiment behind the public pairs, one column per condition",
    "contentUrl": "data/experiments_public.jsonl",
    "encodingFormat": "application/jsonlines",
-   "sha256": "c414a8ad7d0740e541bc462923177ac9630aabdf9dd0bc468602cbcc07a5dfc0"
+   "sha256": "e2f51bd80788a26cad4729309badc9079602d55934885909314f27121f4e89c4"
   },
   {
    "@type": "cr:FileObject",