TA-OPD-10K

English | 中文

TA-OPD-10K is the training dataset released with the paper “Exploring how large language models can empower sequence-based omics tasks”.

10,069 training examples for tool-augmented on-policy distillation (TA-OPD) across six omics tasks. Each example pairs a sequence-based question with an agent answer containing tool-derived biological evidence.

Student and teacher inputs

The student and teacher answer exactly the same question and share the same task instructions. Only the teacher receives the privileged information: the agent answer with its supporting evidence.

• Student: task instructions + question
• Teacher: task instructions + privileged information + the same question

Fields

Field	Content
task	Task identifier
question	The shared question, including the biological sequence
privileged_information	Complete agent answer with tool-derived evidence
student_prompt	Full student messages: system instructions and user question
teacher_prompt	Full teacher messages: the same system instructions, with privileged information added to the user message
ground_truth	Target label

Both prompt fields are ready-to-use lists of {"role": ..., "content": ...} messages.

Quick start

from datasets import load_dataset

train = load_dataset("yj12869741/TA-OPD-10K", split="train")
example = train[0]

student_messages = example["student_prompt"]
teacher_messages = example["teacher_prompt"]
ground_truth = example["ground_truth"]

Apply the model's chat template to these messages. During OPSD training, the student generates its response from student_prompt; the teacher scores the same student-generated response conditioned on teacher_prompt. Only the student is updated.

Using the data with verl

The following adapter targets the teacher-context implementation used in our verl training code. Convert the JSONL dataset to the input format expected by that implementation:

from datasets import load_dataset

data = load_dataset("yj12869741/TA-OPD-10K", split="train")

def to_verl(row, index):
    return {
        "data_source": row["task"],
        "prompt": row["student_prompt"],
        "reward_model": {
            "style": "rule",
            "ground_truth": row["ground_truth"],
            "privileged_solution": row["privileged_information"],
        },
        "extra_info": {"index": index},
    }

data.map(
    to_verl, with_indices=True, remove_columns=data.column_names
).to_parquet("train.parquet")

Apply these settings to the existing OPD training configuration:

data:
  train_files: [train.parquet]
  prompt_key: prompt
  return_raw_chat: true

distillation:
  enabled: true
  teacher_context:
    enabled: true
    ground_truth_key: privileged_solution
    template: null
  distillation_loss:
    loss_mode: k1
    use_policy_gradient: true
    use_task_rewards: false

Here, ground_truth_key selects the privileged information, while reward_model.ground_truth retains the target label. With the default teacher template (template: null), the teacher input is reconstructed from prompt and privileged_solution to match the exported teacher_prompt.

Initialize the student and frozen teacher from the same model checkpoint for OPSD. The student generates responses using only prompt; the teacher scores those same responses with the additional context. Only the student is updated.

The teacher_context settings require our teacher-context extension; they are not a standalone launch configuration for an arbitrary verl version. Model paths, validation data, resource allocation, and optimizer settings remain part of the training recipe.

Tasks

Task	Records
EMP · Epigenetic marks	2,152
Prom · Promoters	1,784
TFBS · Transcription-factor binding sites	1,670
Mod · RNA modifications	1,346
ncRNA · Noncoding RNA families	1,678
EC · Enzymatic functions	1,439
Total	10,069

There are 8,686 unique questions; training resampling is preserved. The data are in English.

File

train.jsonl contains one example per line, with the six fields described above.

Project: OmicsBench.