Datasets:

arjhinety
/

small-mind-post-training-data

Error code:   StreamingRowsError
Exception:    CastError
Message:      Couldn't cast
probe_id: string
persona_id: string
question: string
gold_answer: string
gold_supporting_memory_ids: list<item: string>
  child 0, item: string
category: string
answerable: bool
acceptable_alternatives: list<item: null>
  child 0, item: null
persona: struct<persona_id: string, name: string, description: string, traits: list<item: string>, fact_sheet (... 108 chars omitted)
  child 0, persona_id: string
  child 1, name: string
  child 2, description: string
  child 3, traits: list<item: string>
      child 0, item: string
  child 4, fact_sheet: list<item: struct<fact_id: string, subject: string, predicate: string, object: string, category: str (... 5 chars omitted)
      child 0, item: struct<fact_id: string, subject: string, predicate: string, object: string, category: string>
          child 0, fact_id: string
          child 1, subject: string
          child 2, predicate: string
          child 3, object: string
          child 4, category: string
sessions: list<item: struct<session_id: string, persona_id: string, turns: list<item: struct<turn_id: string,  (... 100 chars omitted)
  child 0, item: struct<session_id: string, persona_id: string, turns: list<item: struct<turn_id: string, session_id: (... 88 chars omitted)
      child 0, session_id: string
      child 1, persona_id: string
      child 2, turns: list<item: struct<turn_id: string, session_id: string, role: string, text: string, ts: int64, reveal (... 33 chars omitted)
          child 0, item: struct<turn_id: string, session_id: string, role: string, text: string, ts: int64, revealed_fact_ids (... 21 chars omitted)
              child 0, turn_id: string
              child 1, session_id: string
              child 2, role: string
              child 3, text: string
              child 4, ts: int64
              child 5, revealed_fact_ids: list<item: string>
                  child 0, item: string
to
{'persona': {'persona_id': Value('string'), 'name': Value('string'), 'description': Value('string'), 'traits': List(Value('string')), 'fact_sheet': List({'fact_id': Value('string'), 'subject': Value('string'), 'predicate': Value('string'), 'object': Value('string'), 'category': Value('string')})}, 'sessions': List({'session_id': Value('string'), 'persona_id': Value('string'), 'turns': List({'turn_id': Value('string'), 'session_id': Value('string'), 'role': Value('string'), 'text': Value('string'), 'ts': Value('int64'), 'revealed_fact_ids': List(Value('string'))})})}
because column names don't match
Traceback:    Traceback (most recent call last):
                File "/src/services/worker/src/worker/utils.py", line 147, 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 127, 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 483, 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 2840, in __iter__
                  for key, example in ex_iterable:
                                      ^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2373, in __iter__
                  for key, pa_table in self._iter_arrow():
                                       ~~~~~~~~~~~~~~~~^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2398, 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 2378, in table_cast
                  return cast_table_to_schema(table, schema)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2306, in cast_table_to_schema
                  raise CastError(
                  ...<3 lines>...
                  )
              datasets.table.CastError: Couldn't cast
              probe_id: string
              persona_id: string
              question: string
              gold_answer: string
              gold_supporting_memory_ids: list<item: string>
                child 0, item: string
              category: string
              answerable: bool
              acceptable_alternatives: list<item: null>
                child 0, item: null
              persona: struct<persona_id: string, name: string, description: string, traits: list<item: string>, fact_sheet (... 108 chars omitted)
                child 0, persona_id: string
                child 1, name: string
                child 2, description: string
                child 3, traits: list<item: string>
                    child 0, item: string
                child 4, fact_sheet: list<item: struct<fact_id: string, subject: string, predicate: string, object: string, category: str (... 5 chars omitted)
                    child 0, item: struct<fact_id: string, subject: string, predicate: string, object: string, category: string>
                        child 0, fact_id: string
                        child 1, subject: string
                        child 2, predicate: string
                        child 3, object: string
                        child 4, category: string
              sessions: list<item: struct<session_id: string, persona_id: string, turns: list<item: struct<turn_id: string,  (... 100 chars omitted)
                child 0, item: struct<session_id: string, persona_id: string, turns: list<item: struct<turn_id: string, session_id: (... 88 chars omitted)
                    child 0, session_id: string
                    child 1, persona_id: string
                    child 2, turns: list<item: struct<turn_id: string, session_id: string, role: string, text: string, ts: int64, reveal (... 33 chars omitted)
                        child 0, item: struct<turn_id: string, session_id: string, role: string, text: string, ts: int64, revealed_fact_ids (... 21 chars omitted)
                            child 0, turn_id: string
                            child 1, session_id: string
                            child 2, role: string
                            child 3, text: string
                            child 4, ts: int64
                            child 5, revealed_fact_ids: list<item: string>
                                child 0, item: string
              to
              {'persona': {'persona_id': Value('string'), 'name': Value('string'), 'description': Value('string'), 'traits': List(Value('string')), 'fact_sheet': List({'fact_id': Value('string'), 'subject': Value('string'), 'predicate': Value('string'), 'object': Value('string'), 'category': Value('string')})}, 'sessions': List({'session_id': Value('string'), 'persona_id': Value('string'), 'turns': List({'turn_id': Value('string'), 'session_id': Value('string'), 'role': Value('string'), 'text': Value('string'), 'ts': Value('int64'), 'revealed_fact_ids': List(Value('string'))})})}
              because column names don't match

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

small-mind-companion — post-training data

Every corpus used to post-train a ~2B vision-language model (google/gemma-4-E2B-it) for long-horizon personalised companion dialogue, in the order it was used: LoRA SFT → LoRA DPO → on-policy distillation.

Part of the OneBee Datasets collection.

Path	Rows	Schema
`sft/v0/{train,val}.jsonl`	202 / 23	`messages`
`sft/v1/{train,val}.jsonl`	2232 / 248	`messages`
`dpo/v0/{train,val}.jsonl`	200 / 23	`prompt`, `chosen`, `rejected`
`dpo/v1_scale/{train,val}.jsonl`	2049 / 228	`prompt`, `chosen`, `rejected`
`distill/v1/{train,val}.jsonl`	2008 / 224	`prompt` only (on-policy)
`training_personas/sft_personas_v0/`	344 probes, 4 personas	probe set + persona JSON
`training_personas/sft_personas_v1/`	3437 probes, 40 personas	probe set + persona JSON
`memory_stores/sft_personas_*/`	4 / 40 SQLite stores	the retrieval corpus each example was built against
`imatrix/calibration.txt`	66,383 lines	GGUF importance-matrix calibration (SFT train + DPO chosen responses)
`imatrix/heldout.txt`	3,806 lines	held-out slice, explicitly not the calibration corpus

messages is a chat-format list. DPO chosen is a real teacher response grounded in injected retrieved memory; rejected is one of five hand-written persona-breaking generic-assistant disclaimers — a deliberately cheap signal encoding "stay in character", not a nuanced quality preference.

The distillation set is prompt-only by design: the student generates its own completions during training and is scored against the teacher's token distribution, so no assistant turns are needed or used.

Regenerating

sft/v1 and dpo/v1_scale are the corpora actually used. sft/v1 was regenerated twice on the day it was created — once to fix a deduplication bug that had silently collapsed ~227 intended abstention examples to 1, and once to rebalance the abstention ratio after that fix over-corrected. sft/v0 and dpo/v0 are the earlier, smaller proof-of-concept runs.

dpo/v1_scale was generated before sft/v1's final regeneration: against the current sft/v1/train.jsonl, 2263/2277 user turns match but 0/2277 system texts do, because the retrieval contexts diverged. The cause was never root-caused. If you chain these corpora, know that the DPO prompts are not byte-identical to the SFT prompts.

Limitations

No human review. Every datasheet says "not human-reviewed", and no reviewer log or teacher transcript was retained, so the teacher identity (gpt-5.6-luna) is self-reported and cannot be verified from any artifact.
Synthetic throughout. Personas, facts and conversations were generated, then the responses were generated against them. This is not recorded human dialogue.
sft/v1's abstention ratio is a tuned parameter, not a principled one. 5% abstention and 6% irrelevant-retrieval were a first attempt; the project documents that a better setting likely exists and was not searched for.
The "by kind" labels in the SFT datasheets are a generation-time annotation, not a field in the committed JSONL, so they cannot be recomputed from these files alone.
Single seed, single run everywhere. Nothing here is replicated.

Provenance

Read from small-mind-companion at the frozen tag study-001 (commit bc80d59, frozen 2026-09-13). Study 001 is frozen: its evidence is hash-pinned in reports/data/study-001-freeze.json and its write-up is at small-mind.arjhinety.com.

Every corpus directory is copied verbatim from the source repository and carries the DATASHEET.md written by its generator plus the hash.txt that pins it. The pins do not all use the same recipe: sft/, dpo/ and distill/ are dataset_hash values (SHA-256 over the concatenated *.jsonl files in sorted order, so DATASHEET.md is not covered by them), while training_personas/sft_personas_v0 and _v1 are benchmark_hash values (a sorted walk over every file, feeding the LF-normalised relative path then the LF-normalised bytes, skipping hash.txt). memory_stores/ and imatrix/ carry no pin in the source repository either. Verify with:

uv run python scripts/recompute_hashes.py

This repository publishes its own corrections. Thirty-four claims the committed artifacts did not support are documented in reports/ERRATA.md, including one that affected this data. Read that before relying on anything here.