ecocoder-scientific-reasoning

ecocoder-scientific-reasoning

Chain-of-Thought (CoT) traces for fine-tuning LLMs on ecological scientific reasoning + code generation.

Each trace follows: [CONTEXT] (ecological problem) → [REASONING] (step-by-step scientific thinking) → [CODE] (runnable R/Python implementation).

Dataset Summary

• 73 unique ecological methods across 18 categories
• Languages: ~60% R, ~40% Python
• Quality-filtered: All traces scored ≥30/100 (median: 85)
• Source: Peer-reviewed papers from PubMed, GBIF Literature, arXiv — distilled via DeepSeek v4 Pro

Method Coverage (18 categories, 73 methods)

Format

Each sample is a chat conversation in ShareGPT/OpenAI format:

{
  "id": "eco-00042",
  "domain": "ecology",
  "method": "maxent",
  "method_category": "sdm",
  "messages": [
    {
      "role": "system",
      "content": "You are an expert ecologist and scientific programmer..."
    },
    {
      "role": "user",
      "content": "We analyzed habitat suitability for Ara macao across..."
    },
    {
      "role": "assistant",
      "content": "[CONTEXT]\n...\n\n[REASONING]\n...\n\n[CODE]\n```r\nlibrary(dismo)\n...\n```"
    }
  ],
  "metadata": {
    "paper_title": "...",
    "paper_id": "arxiv:2405.xxxxx",
    "doi": "10.xxxx/...",
    "language": "R",
    "tags": ["sdm", "maxent"],
    "quality_score": 85.0,
    "source": "deepseek-v4-pro",
    "generated_at": "2026-06-01"
  }
}

Usage

from datasets import load_dataset

ds = load_dataset("alrobles/ecocoder-scientific-reasoning")
print(ds["train"][0]["messages"])

Fine-tuning with Unsloth (Nemotron / Qwen)

from unsloth import FastLanguageModel

model, tokenizer = FastLanguageModel.from_pretrained(
    "nvidia/Nemotron-3-Nano-30B-A3B-BF16",
    max_seq_length=512,
    load_in_4bit=True,
)

# Dataset is already in chat template format — ready for SFTTrainer
from trl import SFTTrainer
trainer = SFTTrainer(
    model=model,
    train_dataset=ds["train"],
    # ...
)

Generation Pipeline

PubMed FTS5 (36M articles)  ──┐
GBIF Literature (61K papers) ──┤
arXiv API                    ──┘
         │
         ▼
    LACS classifier (PU-learning) → domain filtering
         │
         ▼
    DeepSeek v4 Pro (reasoning_effort=high, temperature=0.2)
         │
         ▼
    CoT traces: [CONTEXT] → [REASONING] → [CODE]
         │
         ▼
    Quality filter (≥30/100) + method classification
         │
         ▼
    ecocoder-scientific-reasoning (this dataset)

Related Resources

• EcoSeek — Agentic system for ecological research: github.com/alrobles/ecoseek
• EcoCoder-7B — Fine-tuned ecological code model: huggingface.co/alrobles/EcoCoder-7B
• ecoseek-litdump — Literature pipeline: github.com/alrobles/ecoseek-litdump
• nemotron-eco-reasoner — Nemotron fine-tuning: github.com/alrobles/nemotron-eco-reasoner

Future Expansion

This dataset is designed to grow beyond ecology:

• Genomics — phylogenomics, population genetics, GWAS
• Climate Science — ERA5 downscaling, climate modeling
• Conservation — IUCN assessments, PVA, spatial planning
• Remote Sensing — classification, change detection

Citation

@dataset{robles2026ecocoder,
  title={ecocoder-scientific-reasoning: Chain-of-Thought Training Data for Ecological AI},
  author={Robles Fernandez, Angel Luis},
  year={2026},
  url={https://huggingface.co/datasets/alrobles/ecocoder-scientific-reasoning},
  note={Generated via DeepSeek v4 Pro CoT distillation from PubMed, GBIF, and arXiv}
}

License

Apache 2.0