Dataset Viewer
Auto-converted to Parquet Duplicate
messages
listlengths
2
2
[ { "role": "user", "content": "You are extracting facts from part of a carbon project document (Verra VCS).\nRead the text below and fill in the 7 fields.\nFIELDS:\n- baseline: The counterfactual/business-as-usual scenario that plays out if the project does not happen (usually a deforestation rate or baselin...
[ { "role": "user", "content": "You are extracting facts from part of a carbon project document (Verra VCS).\nRead the text below and fill in the 7 fields.\nFIELDS:\n- baseline: The counterfactual/business-as-usual scenario that plays out if the project does not happen (usually a deforestation rate or baselin...
[ { "role": "user", "content": "You are extracting facts from part of a carbon project document (Verra VCS).\nRead the text below and fill in the 7 fields.\nFIELDS:\n- baseline: The counterfactual/business-as-usual scenario that plays out if the project does not happen (usually a deforestation rate or baselin...
[ { "role": "user", "content": "You are extracting facts from part of a carbon project document (Verra VCS).\nRead the text below and fill in the 7 fields.\nFIELDS:\n- baseline: The counterfactual/business-as-usual scenario that plays out if the project does not happen (usually a deforestation rate or baselin...
[ { "role": "user", "content": "You are extracting one fact from part of a carbon project document (Verra VCS).\nRead the text below and fill in the crediting_period field.\nFIELD:\n- crediting_period: The period in which the project earns carbon credits, with a start and end date, often measured in years\nRU...
[ { "role": "user", "content": "You are extracting one fact from part of a carbon project document (Verra VCS).\nRead the text below and fill in the crediting_period field.\nFIELD:\n- crediting_period: The period in which the project earns carbon credits, with a start and end date, often measured in years\nRU...
[ { "role": "user", "content": "You are extracting facts from part of a carbon project document (Verra VCS).\nRead the text below and fill in the 7 fields.\nFIELDS:\n- baseline: The counterfactual/business-as-usual scenario that plays out if the project does not happen (usually a deforestation rate or baselin...
[ { "role": "user", "content": "You are extracting facts from part of a carbon project document (Verra VCS).\nRead the text below and fill in the 7 fields.\nFIELDS:\n- baseline: The counterfactual/business-as-usual scenario that plays out if the project does not happen (usually a deforestation rate or baselin...
[ { "role": "user", "content": "You are extracting facts from part of a carbon project document (Verra VCS).\nRead the text below and fill in the 7 fields.\nFIELDS:\n- baseline: The counterfactual/business-as-usual scenario that plays out if the project does not happen (usually a deforestation rate or baselin...
[{"role":"user","content":"You are extracting facts from part of a carbon project document (Verra VC(...TRUNCATED)
End of preview. Expand in Data Studio

Tally SFT data (v7 and v8)

Chat-format supervised fine-tuning data for Tally, a quote-grounded Verra carbon-credit field extractor fine-tuned from CohereLabs/tiny-aya-global (Lexsi AI Safety Hackathon, Track 1). Code, labels, labelling rules and the full pipeline: github.com/davissekai/tally-lexsi. Adapter trained on v7: renhuang001/tally-v7-lora. Logs: renhuang001/tally-eval-logs.

Each example is {"messages": [user, assistant]}. The user turn is the extraction instruction plus one page-marked chunk of a Verra project document; the assistant turn is JSON with a value, page and verbatim quote per field, or "not stated".

Config Train Validation Labels Used for
v7 893 (44 projects) 157 Claude Opus teacher labels, quote-checked (v6/v7 rounds) plus carried-over v5 labels Tally v7 adapter (AlignTune TRLSFTTrainer, QLoRA, Kaggle T4)
v8 3206 396 Gemini 3.8 Flash teacher labels with the tally_v8 package, machine-validated (exact-substring quotes, schema, conflict codes) Tally v8 adapter (HF Jobs A100, bf16; adapter not yet public)

Files

  • v7/train.jsonl, v7/validation.jsonl: exactly what the v7 run trained and validated on (messages only).
  • v7/train_with_metadata.jsonl, v7/val_sft_with_metadata.jsonl: same rows plus doc, project, project_group, chunk_index, pages, example_kind and label provenance.
  • v7/eval_miro_bandai*.jsonl: the reserved document-level evaluation set (Miro 2410, Bandai Hills 2929; 23 chunks), never trained on.
  • v8/train.jsonl, v8/validation.jsonl (+ train_with_metadata.jsonl, val_sft_with_metadata.jsonl): the v8 set, including section-aware chunks, single-field questions and conflict-pick examples.
  • v7/stats.json, v8/stats.json: build statistics and split checks.

Example kinds: full 7-field extraction (positive and empty chunks), single-field questions (stated and not-stated), and in v8 conflict-pick examples that ask the model to choose between look-alike candidate values.

Splits and leakage controls

  • Train/validation are split by project group, so no project appears in both.
  • The sealed held-out test projects 5172, 5432, 4329 and neighbour 2928 are excluded from every file here, as is their answer key. Miro (2410) and Bandai (2929) appear only in the v7 reserved eval files.

Source documents and rights

All chunk text comes from VCS project documents (project descriptions, monitoring and verification/validation reports) that are publicly posted on the Verra registry (registry.verra.org). Some copies were downloaded from public mirrors; exact URLs and SHA-256 hashes are in data/sources/MANIFEST.csv and MANIFEST_wide.csv in the GitHub repo. Rights to the source text remain with the document owners (project proponents, validation/verification bodies and Verra). The text is included here only for research and reproducibility; the labels, prompts and structure are the contribution of this dataset (CC-BY-NC-4.0, matching the base model's non-commercial terms). If you are a rights holder and want a document removed, open a discussion on this repo.

Contact e-mail addresses and phone numbers that appear in the source documents were replaced with [email redacted] / [phone redacted] (applied identically to chunks and quotes, so every quote is still an exact substring). Otherwise the text is exactly as trained.

Projects (Verra VCS IDs)

Non-numeric groups (e.g. 737_899_993, 977_981) are grouped neighbouring projects kept on one side of the split.

Documents

Document stems per split
  • v7 train (82): 1067_PD_2012, 1141_PD_2014, 1168_MR_n, 1168_PD, 1168_VR_2013, 1201_PD, 1202_MR_2025, 1202_PD, 1202_VR_2024, 1215_MR_2013, 1215_PD, 1215_VR_2013, 1311_MR_2011, 1311_PD, 1311_VR_2013, 1382_PD_2015, 1408_MR_2022, 1408_PD, 1408_VR_2022, 142_MR_2008, 142_PD_2009, 1463_PD_2015, 1477_PD_2016, 1566_PD_2017, 1650_PD_2014, 1748_PD_2018, 1847_PD_2019, 1855_PD_2019, 1865_PD_2019, 1897_MR_2026, 1897_PD, 1897_VR_2024, 1899_PD_2019, 1900_MR_2026, 1900_PD, 1900_VR_2024, 2070_PD_2019, 2250_PD_n, 2252_PD_2021, 2278_PD_2021, 2297_PD_2024, 2404_PD_2022, 2502_PD_n, 2510_MR_2023, 2510_PD, 2510_VR_2023, 2532_PD_2022, 2551_PD_2022, 2609_PD_2023, 3350_PD_2022, 607_MR_2023, 607_PD_2019, 612_Kasigau-II-Community-Ranches_MR-M9_2022, 612_Kasigau-II-Community-Ranches_PD_excerpt, 612_Kasigau-II-Community-Ranches_Verification_2024, 673_MR_2023, 673_PD, 673_VR_2023, 674_MR_2019, 674_PD_2011, 832_MR_2023, 832_PD_n, 844_MR_2019, 844_PD_2012, 868_MR_2023, 868_PD_2014, 875_MR_2022, 875_PD_2025, 902_MR_2022, 902_PD_2013, 934_Mai-Ndombe_MR-M4_2021, 944_MR_2020, 944_PD_2015, 959_PD_2021, 960_PD_2023, 961_PD_2012, 962_PD_2022, 985_MR_2022, 985_PD_2012, 987_MR_2023, 987_PD, 987_VR_2024
  • v7 validation (18): 1055_PD_2013, 1340_Bale-Mountains_MR_2020-2021, 1622_PD_2017, 1740_PD_n, 1775_MR_2024, 1775_PD, 1775_VR_2023, 562_Kasigau-I-Rukinga_PDD_2011-01-31, 562_Kasigau-I-Rukinga_PD_short_2011-01-25, 737_MR_2016, 737_PD_2011, 899_MR_n, 899_PD_n, 977_PD_2014, 981_PD_2013, 993_MR_n, 993_PD, 993_VR_n
  • v7 reserved eval: 2410_Miro-Ghana_Verification_2020-2022, 2929_Bandai-Hills-Ghana_Validation_2024
  • v8 train (83): 1067_PD_2012, 1168_MR_n, 1168_PD, 1168_VR_2013, 1201_PD, 1215_MR_2013, 1215_PD, 1215_VR_2013, 1311_MR_2011, 1311_PD, 1311_VR_2013, 1340_Bale-Mountains_MR_2020-2021, 1408_MR_2022, 1408_PD, 1408_VR_2022, 142_MR_2008, 142_PD_2009, 1463_PD_2015, 1477_PD_2016, 1566_PD_2017, 1622_PD_2017, 1650_PD_2014, 1740_PD_n, 1748_PD_2018, 1775_MR_2024, 1775_PD, 1775_VR_2023, 1847_PD_2019, 1855_PD_2019, 1865_PD_2019, 1897_MR_2026, 1897_PD, 1897_VR_2024, 1900_MR_2026, 1900_PD, 1900_VR_2024, 2070_PD_2019, 2250_PD_n, 2252_PD_2021, 2278_PD_2021, 2297_PD_2024, 2404_PD_2022, 2502_PD_n, 2510_MR_2023, 2510_PD, 2510_VR_2023, 2532_PD_2022, 2551_PD_2022, 2609_PD_2023, 3350_PD_2022, 562_Kasigau-I-Rukinga_PDD_2011-01-31, 562_Kasigau-I-Rukinga_PD_short_2011-01-25, 607_MR_2023, 607_PD_2019, 612_Kasigau-II-Community-Ranches_MR-M9_2022, 612_Kasigau-II-Community-Ranches_Verification_2024, 674_MR_2019, 674_PD_2011, 737_MR_2016, 737_PD_2011, 832_MR_2023, 832_PD_n, 844_MR_2019, 844_PD_2012, 868_MR_2023, 868_PD_2014, 875_MR_2022, 875_PD_2025, 899_MR_n, 899_PD_n, 902_MR_2022, 902_PD_2013, 934_Mai-Ndombe_MR-M4_2021, 944_MR_2020, 944_PD_2015, 985_MR_2022, 985_PD_2012, 987_MR_2023, 987_PD, 987_VR_2024, 993_MR_n, 993_PD, 993_VR_n
  • v8 validation (16): 1055_PD_2013, 1141_PD_2014, 1202_MR_2025, 1202_PD, 1202_VR_2024, 1382_PD_2015, 1899_PD_2019, 673_MR_2023, 673_PD, 673_VR_2023, 959_PD_2021, 960_PD_2023, 961_PD_2012, 962_PD_2022, 977_PD_2014, 981_PD_2013

Stem format: <VCS project id>_<doc type: PD | MR | VR | ...>_<year>.

Limitations

Teacher labels were machine-verified (quote must be an exact substring of the chunk, unit/type checks, schema) but not fully human-reviewed. v8 labels came from Gemini 3.8 Flash, a different teacher than v7 (Claude Opus). Documents are English-language AFOLU projects only.

Downloads last month
59

Models trained or fine-tuned on renhuang001/tally-sft-data

Collection including renhuang001/tally-sft-data