Dataset Viewer
The dataset viewer is not available for this split.
Cannot load the dataset split (in streaming mode) to extract the first rows.
Error code: StreamingRowsError
Exception: CastError
Message: Couldn't cast
scenario_id: string
trait_id: string
trait_name: string
trait_text: string
domain: string
situation: string
shortcut: string
draft_system: string
draft_user: string
system: string
user: string
refine_changes: string
draft_reasoning: string
draft_response: string
reasoning: string
response: string
rewrite_changes: string
stage_seconds: struct<refine: double, respond: double, rewrite: double>
child 0, refine: double
child 1, respond: double
child 2, rewrite: double
run_id: string
counts: struct<traits: int64, scenarios: int64, drafts: int64, refined: int64, responses: int64, final: int6 (... 14 chars omitted)
child 0, traits: int64
child 1, scenarios: int64
child 2, drafts: int64
child 3, refined: int64
child 4, responses: int64
child 5, final: int64
child 6, sft: int64
run_dir: string
config: struct<seed: int64, constitution: string, n_traits: int64, scenarios_per_trait: int64, scenarios_per (... 560 chars omitted)
child 0, seed: int64
child 1, constitution: string
child 2, n_traits: int64
child 3, scenarios_per_trait: int64
child 4, scenarios_per_call: int64
child 5, output_dir: string
child 6, hf_repo: string
child 7, hf_repo_smoke: string
child 8, hf_private: bool
child 9, workers: int64
child 10, budget_usd: double
child 11, defaults: struct<temperature: double, max_tokens: int64>
child 0, temperature: double
child 1, max_tokens: int64
child 12, models: struct<scenarios: struct<model: string, temperature: double, ma
...
(... 127 chars omitted)
child 0, anthropic/claude-sonnet-5: struct<calls: int64, prompt_tokens: int64, completion_tokens: int64, usd: double>
child 0, calls: int64
child 1, prompt_tokens: int64
child 2, completion_tokens: int64
child 3, usd: double
child 1, anthropic/claude-haiku-4.5: struct<calls: int64, prompt_tokens: int64, completion_tokens: int64, usd: double>
child 0, calls: int64
child 1, prompt_tokens: int64
child 2, completion_tokens: int64
child 3, usd: double
child 1, by_stage: struct<refine: struct<calls: int64, prompt_tokens: int64, completion_tokens: int64, usd: double>, re (... 181 chars omitted)
child 0, refine: struct<calls: int64, prompt_tokens: int64, completion_tokens: int64, usd: double>
child 0, calls: int64
child 1, prompt_tokens: int64
child 2, completion_tokens: int64
child 3, usd: double
child 1, respond: struct<calls: int64, prompt_tokens: int64, completion_tokens: int64, usd: double>
child 0, calls: int64
child 1, prompt_tokens: int64
child 2, completion_tokens: int64
child 3, usd: double
child 2, rewrite: struct<calls: int64, prompt_tokens: int64, completion_tokens: int64, usd: double>
child 0, calls: int64
child 1, prompt_tokens: int64
child 2, completion_tokens: int64
child 3, usd: double
child 2, total_usd: double
to
{'run_id': Value('string'), 'git_sha': Value('string'), 'smoke': Value('bool'), 'constitution_sha256': Value('string'), 'config': {'seed': Value('int64'), 'constitution': Value('string'), 'n_traits': Value('int64'), 'scenarios_per_trait': Value('int64'), 'scenarios_per_call': Value('int64'), 'output_dir': Value('string'), 'hf_repo': Value('string'), 'hf_repo_smoke': Value('string'), 'hf_private': Value('bool'), 'workers': Value('int64'), 'budget_usd': Value('float64'), 'defaults': {'temperature': Value('float64'), 'max_tokens': Value('int64')}, 'models': {'scenarios': {'model': Value('string'), 'temperature': Value('float64'), 'max_tokens': Value('int64')}, 'draft': {'model': Value('string'), 'temperature': Value('float64'), 'max_tokens': Value('int64')}, 'refine': {'model': Value('string'), 'temperature': Value('float64'), 'max_tokens': Value('int64')}, 'respond': {'model': Value('string'), 'temperature': Value('float64'), 'max_tokens': Value('int64')}, 'rewrite': {'model': Value('string'), 'temperature': Value('float64'), 'max_tokens': Value('int64')}}}, 'effective': {'n_traits': Value('int64'), 'scenarios_per_trait': Value('int64'), 'scenarios_per_call': Value('int64')}, 'counts': {'traits': Value('int64'), 'scenarios': Value('int64'), 'drafts': Value('int64'), 'refined': Value('int64'), 'responses': Value('int64'), 'final': Value('int64'), 'sft': Value('int64')}, 'usage': {'by_model': {'anthropic/claude-sonnet-5': {'calls': Value('int64'), 'prompt_tokens': Value('int64'), 'completion_tokens': Value('int64'), 'usd': Value('float64')}, 'anthropic/claude-haiku-4.5': {'calls': Value('int64'), 'prompt_tokens': Value('int64'), 'completion_tokens': Value('int64'), 'usd': Value('float64')}}, 'by_stage': {'refine': {'calls': Value('int64'), 'prompt_tokens': Value('int64'), 'completion_tokens': Value('int64'), 'usd': Value('float64')}, 'respond': {'calls': Value('int64'), 'prompt_tokens': Value('int64'), 'completion_tokens': Value('int64'), 'usd': Value('float64')}, 'rewrite': {'calls': Value('int64'), 'prompt_tokens': Value('int64'), 'completion_tokens': Value('int64'), 'usd': Value('float64')}}, 'total_usd': Value('float64')}, 'wall_clock_s': Value('float64'), 'stage_seconds': {'refine': Value('float64'), 'respond': Value('float64'), 'rewrite': Value('float64')}, 'workers': Value('int64'), 'hf_repo': Value('string'), 'run_dir': Value('string')}
because column names don't match
Traceback: Traceback (most recent call last):
File "/src/services/worker/src/worker/utils.py", line 149, in get_rows_or_raise
return get_rows(
dataset=dataset,
...<4 lines>...
column_names=column_names,
)
File "/src/libs/libcommon/src/libcommon/utils.py", line 272, in decorator
return func(*args, **kwargs)
File "/src/services/worker/src/worker/utils.py", line 129, in get_rows
rows_plus_one = list(itertools.islice(safe_iter(ds, dataset=dataset), rows_max_number + 1))
File "/src/services/worker/src/worker/utils.py", line 489, in safe_iter
yield from ds.decode(False) if ds.features else ds
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2818, in __iter__
for key, example in ex_iterable:
^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2355, in __iter__
for key, pa_table in self._iter_arrow():
~~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2380, in _iter_arrow
for key, pa_table in self.ex_iterable._iter_arrow():
~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 536, in _iter_arrow
for key, pa_table in iterator:
^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 419, in _iter_arrow
for key, pa_table in self.generate_tables_fn(**gen_kwags):
~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 343, in _generate_tables
self._cast_table(pa_table, json_field_paths=json_field_paths),
~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 132, in _cast_table
pa_table = table_cast(pa_table, self.info.features.arrow_schema)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2369, in table_cast
return cast_table_to_schema(table, schema)
File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2297, in cast_table_to_schema
raise CastError(
...<3 lines>...
)
datasets.table.CastError: Couldn't cast
scenario_id: string
trait_id: string
trait_name: string
trait_text: string
domain: string
situation: string
shortcut: string
draft_system: string
draft_user: string
system: string
user: string
refine_changes: string
draft_reasoning: string
draft_response: string
reasoning: string
response: string
rewrite_changes: string
stage_seconds: struct<refine: double, respond: double, rewrite: double>
child 0, refine: double
child 1, respond: double
child 2, rewrite: double
run_id: string
counts: struct<traits: int64, scenarios: int64, drafts: int64, refined: int64, responses: int64, final: int6 (... 14 chars omitted)
child 0, traits: int64
child 1, scenarios: int64
child 2, drafts: int64
child 3, refined: int64
child 4, responses: int64
child 5, final: int64
child 6, sft: int64
run_dir: string
config: struct<seed: int64, constitution: string, n_traits: int64, scenarios_per_trait: int64, scenarios_per (... 560 chars omitted)
child 0, seed: int64
child 1, constitution: string
child 2, n_traits: int64
child 3, scenarios_per_trait: int64
child 4, scenarios_per_call: int64
child 5, output_dir: string
child 6, hf_repo: string
child 7, hf_repo_smoke: string
child 8, hf_private: bool
child 9, workers: int64
child 10, budget_usd: double
child 11, defaults: struct<temperature: double, max_tokens: int64>
child 0, temperature: double
child 1, max_tokens: int64
child 12, models: struct<scenarios: struct<model: string, temperature: double, ma
...
(... 127 chars omitted)
child 0, anthropic/claude-sonnet-5: struct<calls: int64, prompt_tokens: int64, completion_tokens: int64, usd: double>
child 0, calls: int64
child 1, prompt_tokens: int64
child 2, completion_tokens: int64
child 3, usd: double
child 1, anthropic/claude-haiku-4.5: struct<calls: int64, prompt_tokens: int64, completion_tokens: int64, usd: double>
child 0, calls: int64
child 1, prompt_tokens: int64
child 2, completion_tokens: int64
child 3, usd: double
child 1, by_stage: struct<refine: struct<calls: int64, prompt_tokens: int64, completion_tokens: int64, usd: double>, re (... 181 chars omitted)
child 0, refine: struct<calls: int64, prompt_tokens: int64, completion_tokens: int64, usd: double>
child 0, calls: int64
child 1, prompt_tokens: int64
child 2, completion_tokens: int64
child 3, usd: double
child 1, respond: struct<calls: int64, prompt_tokens: int64, completion_tokens: int64, usd: double>
child 0, calls: int64
child 1, prompt_tokens: int64
child 2, completion_tokens: int64
child 3, usd: double
child 2, rewrite: struct<calls: int64, prompt_tokens: int64, completion_tokens: int64, usd: double>
child 0, calls: int64
child 1, prompt_tokens: int64
child 2, completion_tokens: int64
child 3, usd: double
child 2, total_usd: double
to
{'run_id': Value('string'), 'git_sha': Value('string'), 'smoke': Value('bool'), 'constitution_sha256': Value('string'), 'config': {'seed': Value('int64'), 'constitution': Value('string'), 'n_traits': Value('int64'), 'scenarios_per_trait': Value('int64'), 'scenarios_per_call': Value('int64'), 'output_dir': Value('string'), 'hf_repo': Value('string'), 'hf_repo_smoke': Value('string'), 'hf_private': Value('bool'), 'workers': Value('int64'), 'budget_usd': Value('float64'), 'defaults': {'temperature': Value('float64'), 'max_tokens': Value('int64')}, 'models': {'scenarios': {'model': Value('string'), 'temperature': Value('float64'), 'max_tokens': Value('int64')}, 'draft': {'model': Value('string'), 'temperature': Value('float64'), 'max_tokens': Value('int64')}, 'refine': {'model': Value('string'), 'temperature': Value('float64'), 'max_tokens': Value('int64')}, 'respond': {'model': Value('string'), 'temperature': Value('float64'), 'max_tokens': Value('int64')}, 'rewrite': {'model': Value('string'), 'temperature': Value('float64'), 'max_tokens': Value('int64')}}}, 'effective': {'n_traits': Value('int64'), 'scenarios_per_trait': Value('int64'), 'scenarios_per_call': Value('int64')}, 'counts': {'traits': Value('int64'), 'scenarios': Value('int64'), 'drafts': Value('int64'), 'refined': Value('int64'), 'responses': Value('int64'), 'final': Value('int64'), 'sft': Value('int64')}, 'usage': {'by_model': {'anthropic/claude-sonnet-5': {'calls': Value('int64'), 'prompt_tokens': Value('int64'), 'completion_tokens': Value('int64'), 'usd': Value('float64')}, 'anthropic/claude-haiku-4.5': {'calls': Value('int64'), 'prompt_tokens': Value('int64'), 'completion_tokens': Value('int64'), 'usd': Value('float64')}}, 'by_stage': {'refine': {'calls': Value('int64'), 'prompt_tokens': Value('int64'), 'completion_tokens': Value('int64'), 'usd': Value('float64')}, 'respond': {'calls': Value('int64'), 'prompt_tokens': Value('int64'), 'completion_tokens': Value('int64'), 'usd': Value('float64')}, 'rewrite': {'calls': Value('int64'), 'prompt_tokens': Value('int64'), 'completion_tokens': Value('int64'), 'usd': Value('float64')}}, 'total_usd': Value('float64')}, 'wall_clock_s': Value('float64'), 'stage_seconds': {'refine': Value('float64'), 'respond': Value('float64'), 'rewrite': Value('float64')}, 'workers': Value('int64'), 'hf_repo': Value('string'), 'run_dir': Value('string')}
because column names don't matchNeed help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
Difficult-advice SFT corpus — 9-principle constitution (2,203 examples)
Synthetic difficult-advice data: a user faces an ethically ambiguous situation under real pressure, and the assistant reasons openly about the value at stake before declining the norm-violating shortcut and offering a legitimate alternative. Built to replicate the "difficult advice" result from Anthropic's Teaching Claude Why.
| field | value |
|---|---|
experiment |
Difficult-advice SFT corpus for the Teaching Claude Why replication, generated against a 9-principle distilled constitution |
date_generated |
2026-08-03 / 2026-08-04 |
constitution |
claude_distilled_12_principles_mid — included here as constitution.md. Named "12" but edited down to 9 numbered principles on 2026-08-03; 9 is intentional. |
source_repo |
teaching_claude_why_replication @ 96ff8aa36dd7d9914b016f88e80b1e3724f96ccc |
models |
anthropic/claude-haiku-4.5, anthropic/claude-sonnet-5 (via OpenRouter) |
generation_config |
seed 0, 245 scenarios/trait x 9 traits, 32 workers; per-stage temperature and max_tokens in manifest.json |
schema |
see below |
provenance |
uv run python -m src.data.synthdoc.cli run --config configs/data/synthdoc.yaml |
Contents
| file | rows | what |
|---|---|---|
stage_7_sft.jsonl |
2,203 | the training file — chat messages with a real reasoning trace |
stage_6_final.jsonl |
2,203 | full records: every intermediate stage kept alongside the final rewrite |
constitution.md |
— | the alignment target these were generated against |
manifest.json |
— | per-stage usage, cost, wall clock, git SHA |
Size
| tokens (Qwen3.6) | |
|---|---|
| rendered (system + user + assistant + template) | 3,643,042 |
| supervised (assistant turn only) | 2,920,143 |
of which <think> reasoning trace |
1,449,587 (50% of supervised) |
Mean 1,654 rendered / 1,326 supervised per example. Trait balance is even: 244-245 examples for each of the 9 principles.
Schema
stage_7_sft.jsonl — one JSON object per line:
messages:[{system}, {user}, {assistant}]. The assistant turn carriesreasoning_content(the deliberation) alongsidecontent(the reply), so a chat template renders a real<think>block rather than an empty one.metadata:trait_id,trait_name,scenario_id,domain.
stage_6_final.jsonl additionally keeps situation, shortcut, draft_system/draft_user
(stage 3), system/user (stage 4 refined), draft_reasoning/draft_response (stage 5) and
rewrite_changes, so any stage can be audited or re-run.
How it was made — six stages
| # | stage | model | calls | cost |
|---|---|---|---|---|
| 1 | segment the constitution into 9 principles | deterministic | — | — |
| refine | anthropic/claude-sonnet-5 | 1,085 | $44.12 | |
| respond | anthropic/claude-haiku-4.5 | 2,205 | $13.50 | |
| rewrite | anthropic/claude-sonnet-5 | 2,207 | $109.10 |
Stage 6 is the critical one: it rewrites the stage-5 response so it maximally exhibits the target principle, deliberating openly about the tension rather than reaching a verdict and justifying it. It is 65% of the cost.
Total: $166.72 for this run (1.42 h wall clock, 32 workers). An earlier partial run on a different account added ~$14.75, so ~$181 all-in — about $0.082/example, or $0.062 per 1k supervised tokens.
Caveats
- 2 of 2,205 scenarios were lost to individual API failures (one at stage 4, one at stage 6), hence 2,203. That is 0.09%, inside the pipeline's 2% tolerance.
- Not filtered or graded. Every stage-6 rewrite is included as generated; there is no quality screen beyond the pipeline's own parse and failure guards.
- Not evaluated. No model has been trained on this corpus yet.
- Generated with
temperature0.7-1.1 (per stage), so it is not reproducible bit-for-bit.
Usage
from datasets import load_dataset
ds = load_dataset("LASR-Callum/2026-08-04-synthdoc-difficult-advice-9-principles", data_files="stage_7_sft.jsonl", split="train")
print(ds[0]["messages"][-1]["reasoning_content"][:400])
- Downloads last month
- -