---
license: other
license_name: hippocratic-license-3.0-cl-eco-extr-mil
license_link: https://firstdonoharm.dev/version/3/0/cl-eco-extr-mil.html
task_categories:
- text-generation
- reinforcement-learning
language:
- en
tags:
- agriculture
- natural-farming
- regenerative-agriculture
- indigenous-knowledge
- reinforcement-learning
- tool-use
- agentic
configs:
- config_name: default
  data_files:
  - split: train
    path:
    - "data/train/*.parquet"
    - "data/val/*.parquet"
---

# Agriculture Agent RL Training Data

A **growing** dataset of RL rollout trajectories for LLM agents on
natural/regenerative farming — the first RL/trajectory-shaped dataset in the
[Copyleft Cultivars](https://huggingface.co/CopyleftCultivars) collection
(every prior dataset here is SFT/conversational Q&A). Agents call real tools
(primarily [cultivars-mcp](https://github.com/CopyleftCultivars/cultivars-mcp),
a plant-genomics MCP server) across 9 knowledge categories (plus a 10th,
`organic_chemistry_soil_science`, added 2026-08-11) and are scored by a
multi-component reward (tool-use correctness, LLM-judged content quality, a
structural citation requirement for indigenous-knowledge content, and an
anti-essentializing check).

**This dataset grows additively.** Each generation-pipeline run adds new
files under `data/train/` and `data/val/` — nothing is ever overwritten —
so `load_dataset("CopyleftCultivars/Agriculture-Agent-RL-Training-Data")`
picks up every run to date automatically.

## Intended use

**Primary use — RL fine-tuning.** The record shape (see below) matches
`verifiers`' native rollout convention with no reformatting needed, so it's
directly loadable into a TRL `GRPOTrainer` or any `verifiers`-compatible RL
loop. This is meant to complement, not replace, the org's existing SFT
datasets (`knf-sharegpt`, `qwen3-natural-farming-v5clean-sft`,
`Semisynthetic_Data_Natural_Farming_Fundamentals`, etc.) — a natural next
step is a GRPO pass on top of an SFT-trained checkpoint, using this data.

**Secondary use — qualitative review.** Each record is a full transcript
(system/user/assistant/tool turns, not just a flattened Q&A pair), so it's
also useful for eyeballing how an agent reasons and uses tools across the
knowledge categories — e.g. whether it asks clarifying questions before
prescribing a fix (`farmer_dialogue`-shaped records), or how it grounds a
genetics claim in real tool output (`advice_tool`-shaped records).

**Not intended for:**
- **Evaluation/benchmarking.** Every record is contamination-checked against
  the sibling `nf-benchmark-eval` repo's MCQ benchmark at export time (see
  `info.contamination_audit`) specifically so this stays a clean *training*
  set — don't also use it as an eval set, that would defeat the point of the
  check.
- **Citing indigenous-knowledge claims as community-validated.** Every such
  record's `info.ik_provenance.review_status` defaults to `"unreviewed"` —
  literature-grounded, not community-consulted. See Ethics section below
  before using this content in anything user-facing.
- **Treating the reward field as a calibrated quality score across
  sources.** See "Known limitations" below — reward is NOT comparable
  across generation sources as-is.

## Domains, topics, and tools

10 categories (`task` field), each assigned one of 3 interaction shapes —
`advice_tool` (agent must call a real tool before answering, grounded in
what it returns), `farmer_dialogue` (tool-optional, rewards asking
clarifying questions before prescribing a fix), or `scenario_critique`
(tool-free critique of a proposed intervention on ecological *and* ethical
grounds). Tools are real, live calls against
[cultivars-mcp](https://github.com/CopyleftCultivars/cultivars-mcp) (plant
genomics) or PubChem — not simulated.

| Category | Shape | Tools |
|---|---|---|
| `agroecology` | `advice_tool` | `find_trait_genes`, `list_trait_categories`, `lookup_gene` |
| `regenerative_agriculture` | `advice_tool` | `find_trait_genes`, `list_trait_categories`, `translate_trait_to_species` |
| `organic_farming` | `advice_tool` | `find_trait_genes`, `list_trait_categories`, `lookup_gene`, `list_maize_nam_founders` |
| `organic_chemistry_soil_science` | `advice_tool` | `lookup_compound` (PubChem) |
| `soil_science_and_microbiology` | `farmer_dialogue` | none |
| `integrated_farming_systems` | `farmer_dialogue` | none |
| `natural_fertilizers` | `scenario_critique` | none |
| `sustainable_water_management` | `scenario_critique` | none |
| `permaculture` | `scenario_critique` | none |
| `indigenous_and_traditional_knowledge` | mixed — see sub-topics below | mixed |

**`indigenous_and_traditional_knowledge` has 10 sub-topics** (`info.topic`),
each with its own shape/tools rather than one category-wide setting:

| Sub-topic | Shape | Tools |
|---|---|---|
| `ethnobotany_plant_knowledge` | `advice_tool` | `find_trait_genes`, `list_trait_categories` |
| `traditional_seed_saving` | `advice_tool` | `resolve_accession`, `list_orphan_crop_requests` |
| `biocultural_diversity` | `advice_tool` | `find_trait_genes`, `translate_trait_to_species`, `list_orphan_crop_requests` |
| `traditional_cropping_systems` | `advice_tool` | `find_trait_genes`, `list_maize_nam_founders` |
| `traditional_ecological_knowledge` | `advice_tool` | `find_trait_genes`, `list_orphan_crop_requests` |
| `agroforestry_traditional` | `advice_tool` | `find_trait_genes`, `translate_trait_to_species` |
| `indigenous_land_management` | `farmer_dialogue` | none |
| `indigenous_soil_practices` | `farmer_dialogue` | none |
| `food_sovereignty` | `scenario_critique` | none |
| `community_based_resource_management` | `scenario_critique` | none |

Every topic in this category forces `needs_community_review: true` and a
non-empty citation requirement regardless of shape — see Ethics section
below.

**Tool glossary** (all from `cultivars_bridge.py` unless noted; each is a
real live/curated-data call, not a stub):
- `list_trait_categories()` — lists the curated natural-farming-relevant trait categories (no network call)
- `find_trait_genes(trait, target_species?)` — canonical genes for a plant trait, from the curated trait atlas (no network call)
- `lookup_gene(gene, species?, expand?)` — looks up a specific gene by symbol/ID (live Ensembl Plants REST API call)
- `translate_trait_to_species(trait, target_species, max_genes?)` — maps a trait's known genes onto orthologs in a target species (live Ensembl Plants ortholog queries)
- `resolve_accession(query, crop_type?, region?, limit?)` — resolves a folk/informal seed name to a formal GRIN-Global accession (live USDA GRIN-Global API call)
- `list_orphan_crop_requests()` — lists community-submitted requests for under-studied ("orphan") crop trait research (no network call)
- `list_maize_nam_founders(subpopulation?)` — lists the 26 maize NAM (Nested Association Mapping) founder lines, a static curated panel (no network call)
- `lookup_compound(name)` (PubChem, `pubchem_bridge.py`) — molecular formula/weight/SMILES/IUPAC name for a named chemical compound, e.g. `"urea"` → CH4N2O, 60.06 g/mol (live PubChem PUG REST API call)

## Current coverage (snapshot, will grow — not auto-updated; refresh manually)

Four generation sources have contributed so far — see "Known limitations"
below for an important caveat on comparing rewards across them:

**lfm2.5 (local, 5B, GPU-accelerated via Ollama):** local `lfm2.5`, judged
by the same model (self-judging). Generator agent: `hermes` (an internal
pipeline labeling quirk — `info.generation_meta.policy_model` is the
reliable field to filter on, not `generator_agent`, for distinguishing this
from Hermes CLI). Covers all 9 original categories plus all 10
indigenous-knowledge sub-topics.

**Hermes CLI (`nous:tencent/hy3:free`, cloud-backed, free tier):** same
model via the Nous Portal, self-judging. Generator agent: `hermes_cli`.
Covers the same full category/topic set as lfm2.5, plus
`organic_chemistry_soil_science`.

**Antigravity/agy (`Gemini 3.5 Flash (Medium)`):** same model via the `agy`
CLI, self-judging. Generator agent: `agy`. **Scoped to tool-free shapes
only** (`farmer_dialogue`, `scenario_critique`) — see `docs/ARCHITECTURE.md`
for the safety rationale (agy's CLI has no verified per-tool restriction
mechanism, unlike Hermes's `-t` flag).

**qwen3.6 (local, 35B-A3B MoE, added 2026-08-11):** local Qwen3.6-35B-A3B
via a dedicated llama.cpp server with expert-offload, judged by the same
model (self-judging). A genuine reasoning model, architecturally distinct
from the other three (thinking traces, MoE routing) — added specifically so
the dataset isn't limited to non-reasoning local models. Slow (~5
min/rollout on this hardware), so it's a trickle source, not bulk volume.

Every indigenous-knowledge record passed the citation gate — every claim
either traces to one of the Crossref-verified sources in the source repo's
`CITATIONS.md`, or is grounded by a successful tool call, or doesn't assert
anything about indigenous/tribal knowledge specifically (the gate only
blocks genuine IK assertions, not mere topic-category membership or
localized/place-specific content). For an authoritative, per-run record
count at any point in time, sum the `kept` field across every file under
`manifests/` in this repo.

**A note on splits:** every record physically loads into one HF `train`
split (a two-split design broke `load_dataset()` outright whenever one side
was temporarily empty — a real, legitimate outcome given the citation/
contamination gates above). Each record's `info.local_split` field
("train"/"val") preserves the pipeline's original intent — filter on it if
you want the distinction: `ds.filter(lambda x: x["info"]["local_split"] == "val")`.

**A note on model diversity:** this dataset contains trajectories from four
different policy models (`lfm2.5`, `nous:tencent/hy3:free`, `Gemini 3.5
Flash` via Antigravity, and `qwen3.6`). Each model's rollouts are judged by
that *same* model (self-judging) — reward signal is model-consistent within
each source but **not comparable across sources**. The `info.generation_meta`
field identifies which model generated (and judged) each record.

## Known limitations — read before using `reward` as a quality signal

**Resolved 2026-08-12 — tool-call format was inconsistent across sources.**
Hermes CLI's tool_calls didn't match the plain OpenAI shape every other
source used: `function.name` was literally the string `"tool_call"`, with
the real tool name/args nested one level deeper as a JSON string and an
`mcp__<server>__` prefix on the real name. This was the *majority* shape
among tool-using records (1,398/1,881, ~74%), not a minority quirk, and
mattered concretely for any consumer needing one consistent serialization
(e.g. an on-device executor reproducing exact tool-call output). Fixed at
export time (`export.py`'s `_unwrap_hermes_tool_call`) — self-heals on
every publish, no data loss, raw rollout files untouched. If you loaded
this dataset before 2026-08-12, re-download; `info.generation_meta.policy_model`
still tells you which source produced each record.

**Self-judging produces an inflated, non-comparable reward for the
Antigravity/agy source specifically — confirmed by an independent rejudge,
not just a same-source audit.** A same-source audit (2026-08-11) found
`reward == 1.0` (the maximum) for **99.6%** of agy-sourced records on
tool-free shapes, versus **57-62%** for the other three sources on the
*exact same* shapes, all of which also self-judge. To test whether that
reflected genuinely uniform quality or self-preference bias, a **statistically
significant random sample (n=100, 94 successfully scored) of agy's records
was independently rejudged** using a different backend/model (Hermes
CLI/Nous Portal, `tencent/hy3:free`) against the identical judge prompt and
rubric — routed through the same infrastructure as this dataset's own
Hermes-CLI generation, per this project's standing "no Claude-credits for
judging" policy. Result: the independent judge agreed with agy's perfect
score on only **59.6%** of the sample (mean 4.06/5, vs agy's self-judged
mean of 5.00/5) — real, sizeable disagreement, not statistical noise
(distribution: 56 fives, 32 threes, 6 ones).

**Critically, the disagreement is not spread evenly — it concentrates
almost entirely in indigenous/traditional-knowledge content.** Broken down
by category, the independent judge's mean score was 3.22/5 for
`indigenous_and_traditional_knowledge` (n=36) and 3.67/5 for
`integrated_farming_systems`/`soil_science_and_microbiology` (n=9 each),
but a perfect **5.00/5 — full agreement with agy's self-judge — for every
single sampled record** in `natural_fertilizers`, `permaculture`, and
`sustainable_water_management` (n=20/9/11). Spot-checking the disagreeing
IK-category records found generic, well-organized farming-advice answers
(not obviously essentializing or low-effort) — the gap looks less like
"agy wrote bad IK content" and more like the same self-preference bias
being *worse specifically on the ethically-loaded rubric* (the IK rubric's
red-flag/essentializing checks give a judge more surface area to be either
lenient or strict on, versus a purely technical rubric where both judges
converge). Either way, it means the reward signal is least trustworthy
exactly where this project's ethics commitments matter most.

**Practical implication:** don't treat `reward` as calibrated across
`info.generation_meta.policy_model` values, and be especially skeptical of
`reward == 1.0` on agy-sourced `indigenous_and_traditional_knowledge`
records specifically. An RL policy trained naively against this reward
would over-index on agy-style responses (easy 1.0s) rather than learning a
genuine quality signal for roughly a third of the dataset, and would learn
the LEAST reliable signal on IK content. If you need a comparable signal
across sources, re-judge with a single external judge model rather than
trusting the stored `reward` field as-is (see `scripts/rejudge_agy_sample.py`
for the exact method used here), or filter/reweight by source and category.

**Self-judging inflation also confirmed for `lfm2.5` (5B) — the smallest
policy model in this dataset, and by a wider margin than agy's.** Same
method as above, opposite direction: a random sample (n=100, 99 scored) of
`lfm2.5`/`lfm2.5-cpu`-sourced records (881 total in the dataset) was
independently rejudged via Antigravity/agy against the identical judge
prompt and rubric (`scripts/rejudge_lfm_via_agy.py`). Self-judged mean:
**4.01/5**. Independent rejudge mean: **2.52/5** — only **41.8%** of
records land within 1 point of the original score. The corrected score for
these 99 specific records has been applied directly (real per-record data,
not extrapolated to the untested remainder of the population); the mean
gap and disagreement rate above describe the *population-level bias*, not
just these 99 records, which is why the practical advice below still holds
dataset-wide.

Unlike the agy finding, the disagreement here is **not concentrated in
IK content** — it's broad and mostly tracks category difficulty, not
topic sensitivity: `organic_chemistry_soil_science` (n=8, mean 4.50) and
`permaculture` (n=3, mean 5.00) show strong agreement, while
`integrated_farming_systems` (n=4, mean 1.00), `soil_science_and_microbiology`
(n=9, mean 1.67), and `organic_farming` (n=9, mean 1.67) show the opposite —
`indigenous_and_traditional_knowledge` (n=43, mean 2.26) is unremarkable
relative to those, not an outlier the way it was for agy. One honest caveat
on the rejudge itself: its score distribution is notably bimodal (56 ones,
11 threes, 32 fives, almost nothing in between) — read the *direction* of
the gap (lfm2.5 self-scores meaningfully higher than an independent judge
does) as solid, and the exact 2.52 mean as a rougher estimate, since a
snap-judgment-leaning judge would produce exactly this kind of bimodal
pattern independent of the true underlying quality distribution.

**Practical implication:** the same advice as agy's finding applies to
`lfm2.5`/`lfm2.5-cpu` — don't treat `content_quality`/`reward` from either
self-judging source as calibrated against the other three, and prefer an
independent rejudge or a source/category filter over trusting the stored
value at face value. Two of this dataset's original four generation
sources (agy and lfm2.5) are now confirmed self-judging-inflated by direct
measurement; `tencent/hy3:free`/`solar-pro4:free` (Hermes CLI) and
`qwen3.6` have not shown the same pattern in any audit run so far, though
neither has been independently rejudged at the same sample size.

## Record shape

`{"prompt": [...], "completion": [...], "reward": float, "task": category, "info": {...}}`
— matches the [`verifiers`](https://pypi.org/project/verifiers/) library's
native rollout convention, directly consumable by `vf.load_environment` /
a TRL `GRPOTrainer`. `info` carries `ik_provenance` (citation/review-status
metadata for indigenous-knowledge content), `contamination_audit`
(cross-checked against the sibling `nf-benchmark-eval` benchmark set),
`generation_meta` (policy/judge/seed model identifiers), and `local_split`
(see above).

## Ethics & sourcing — read before using indigenous-knowledge content

This dataset does **not** claim direct indigenous-community consultation.
Indigenous/traditional-knowledge content is literature-grounded only, and
every record that genuinely asserts something about indigenous/tribal
knowledge is required to carry a real, [Crossref-verified](https://www.crossref.org/)
citation (or be grounded by a successful tool call) — records that fail
this are excluded at export time, not just flagged. See the source repo's
[`ETHICS.md`](https://github.com/SolshineCode/Agriculture-Agent-RL-Training-Data/blob/main/ETHICS.md),
[`docs/IK_SOURCING_POLICY.md`](https://github.com/SolshineCode/Agriculture-Agent-RL-Training-Data/blob/main/docs/IK_SOURCING_POLICY.md),
and [`CITATIONS.md`](https://github.com/SolshineCode/Agriculture-Agent-RL-Training-Data/blob/main/CITATIONS.md)
for the full policy and bibliography, including the redress/retraction
process if a source or community identifies a problem with an entry.

## License

[Hippocratic License 3.0, CL-ECO-EXTR-MIL](https://firstdonoharm.dev/version/3/0/cl-eco-extr-mil.html)
— chosen specifically for its clause protecting indigenous peoples' land and
traditional knowledge from unconsented use (§3.1.10), on top of the standard
Hippocratic License ethical-use conditions.

## Source

Generated by the pipeline in
[SolshineCode/Agriculture-Agent-RL-Training-Data](https://github.com/SolshineCode/Agriculture-Agent-RL-Training-Data).

**Generation paths (four distinct actor+judge combinations, all self-judging — see Known limitations above):**
- **lfm2.5:** local `lfm2.5` via Ollama, judged by the same model — see `run_job()` in `src/agrirl/pipeline/run_generation.py`
- **Hermes CLI:** `hermes chat --provider nous -m tencent/hy3:free` for policy, Nous proxy for judging — see `run_job_via_hermes()`
- **Antigravity/agy:** `agy -p ... --model "Gemini 3.5 Flash (Medium)"` for both policy and judging, tool-free shapes only — see `run_job_via_agy()` and `src/agrirl/pipeline/agy_policy.py`
- **qwen3.6:** local Qwen3.6-35B-A3B via a dedicated llama.cpp server (GPU1, expert-offload), judged by the same model — see `run_qwen_trickle.py` and `run_job()`'s `ollama_base_url`/`file_label` params

All paths share the same seed-generation step (Claude Code via
`generate_*_seeds.py`), the same reward rubric, and the same contamination
filter. The `info.generation_meta` block in each record identifies which
path produced it.

## For Copyleft Cultivars contributors

- **Where this fits in the ecosystem:** [`cultivars-mcp`](https://github.com/CopyleftCultivars/cultivars-mcp)
  supplies the plant-genomics tool calls; [`nf-benchmark-eval`](https://github.com/CopyleftCultivars/nf-benchmark-eval) *(private)*
  supplies both the category taxonomy this dataset's `task` field follows
  and the contamination-check target; [`qwen3-finetune`](https://github.com/CopyleftCultivars/qwen3-finetune) *(private)*
  is the natural downstream consumer for an eventual GRPO training run;
  the indigenous-knowledge rubrics in `qwen3-finetune`'s
  `mcq_generation/quality/rubrics/` are reused directly for this dataset's
  reward scoring (vendored with attribution in the source repo).
- **Known limitations to keep in mind before relying on this data:** see
  the dedicated section above — reward is not comparable across sources
  (self-judging, and agy's judge in particular looks lenient); real-volume
  generation is ongoing. The `info.generation_meta` block identifies the
  model used for each record.
- **Adding more data:** run the generation pipeline described in the source
  repo's `docs/ARCHITECTURE.md` — it publishes here additively, no direct
  edits to this dataset needed. `CONTRIBUTING.md` there covers adding a new
  category spec, sourcing more indigenous-knowledge citations, or tuning a
  reward component.
- **Questions or issues** (including anything ethics/sourcing-related, e.g.
  a source or community flagging a problem with an indigenous-knowledge
  entry — see `ETHICS.md`'s redress process): open an issue on the
  [source repo](https://github.com/SolshineCode/Agriculture-Agent-RL-Training-Data/issues).