Add learn_to_ask example

examples/learn_to_ask/README.md

@@ -0,0 +1,143 @@
+# Learn2Ask: Getting Started
+
+This guide demonstrates how to train a proactive LLM using the **Learn2Ask** framework from [Grounded in Reality: Learning and Deploying Proactive LLM from Offline Logs](https://arxiv.org/abs/2510.25441).
+**Hardware requirement**: ≥32 H20 (or equivalent) GPUs for full-scale reproduction.
+
+All relevant files are located under `examples/learn_to_ask/`:
+- Workflow & prompts: `examples/learn_to_ask/workflow/`
+- Training config: `examples/learn_to_ask/train.yaml`
+- Data preparation scripts: `examples/learn_to_ask/data_prepare/`
+
+---
+
+## Step 1. Prepare Datasets
+
+Download the [RealMedConv](https://huggingface.co/datasets/datajuicer/RealMedConv) dataset (`.jsonl` format). Each line is a full conversation log:
+
+```json
+{
+  "session_id": 35310,
+  "diagn": "Upper Respiratory Tract Infection",
+  "messages": [{"role": "user", "content": "Sore throat, phlegm, red eyes, cough, hoarse voice"}, {"role": "user", "content": "I took Amoxicillin"}, {"role": "user", "content": "But I still don't feel well"}, {"role": "user", "content": "Mainly it's a respiratory infection, sore throat, phlegm, hoarse voice, red eyes"}, {"role": "user", "content": "When I wake up, there is a lot of eye discharge, and a lot of phlegm"}, {"role": "assistant", "content": "How long have the symptoms been present?"}, {"role": "user", "content": "About 2 days"}, {"role": "user", "content": "My eyes are very red"}, {"role": "assistant", "content": "Is there any discharge?"}, {"role": "user", "content": "Yes"}, {"role": "user", "content": "Please check my description, I wrote all the details"}, {"role": "assistant", "content": "Sure"}, {"role": "assistant", "content": "The internet was down just now"}, {"role": "user", "content": "Okay"}, {"role": "assistant", "content": "Is the discharge thick, thin, or stringy?"}, {"role": "user", "content": "It's thick"}, {"role": "user", "content": "Yellowish"}, {"role": "user", "content": "Mainly a lot in the morning, and a lot of phlegm"}, {"role": "assistant", "content": "Does it affect your vision? Do you have eye pain? Itchy eyes? Foreign body sensation? Tears?"}, {"role": "user", "content": "No"}, {"role": "user", "content": "Mainly still sore throat"}, {"role": "user", "content": "The eyes are just red and have discharge"}, {"role": "user", "content": "Sore throat, a lot of phlegm, mild cough, hoarse voice"}, {"role": "assistant", "content": "Okay"}, {"role": "assistant", "content": "Have you had any medical examinations or medication history? Any history of drug allergies or chronic diseases?"}, {"role": "user", "content": "No"}, {"role": "user", "content": "Please help as soon as possible, it's getting late"}, {"role": "assistant", "content": "<med_search>"}]
+}
+```
+You need to perform the following preprocessing steps to turn the log in to training/testing samples for our `learn_to_ask` framework, there are two simple steps:
+- Segment the original conversation log (session) into context–future pairs, then extract `info_truth` labels from the `remaining_chat` field.
+```bash
+python examples/learn_to_ask/data_prepare/1_info_extract_pipeline.py --input_file /path/to/RealMedConv/train.jsonl --output_file examples/learn_to_ask/data_raw/train_processed.jsonl
+```
+
+- Convert these samples into final training/testing datasets.
+```bash
+python examples/learn_to_ask/data_prepare/2_build_dataset.py --input_file examples/learn_to_ask/data_raw/train_processed.jsonl --output_file examples/learn_to_ask/data/train.jsonl
+```
+
+These scripts are implementations of the following procedures.
+
+### Segment into Context–Future Pairs
+For each turn in a session, split the conversation into:
+- `messages`: the **observed context** up to that point,
+- `remaining_chat`: the **subsequent dialogue** (i.e., the "future" from that turn onward).
+Each segmented sample should include a unique `cid` (e.g., `{session_id}_{turn_index}`).
+```JSON
+{
+   "cid": "35310_7",
+   "session_id": "35310",
+   "diagn": "Upper Respiratory Tract Infection",
+   "messages": [{"role": "user", "content": "Sore throat, phlegm, red eyes, cough, hoarse voice"}, {"role": "user", "content": "I took Amoxicillin"}, {"role": "user", "content": "But I still don't feel well"}, {"role": "user", "content": "Mainly it's a respiratory infection, sore throat, phlegm, hoarse voice, red eyes"}, {"role": "user", "content": "When I wake up, there is a lot of eye discharge, and a lot of phlegm"}, {"role": "assistant", "content": "How long have the symptoms been present?"}, {"role": "user", "content": "About 2 days"}, {"role": "user", "content": "My eyes are very red"}],
+   "remaining_chat": [{"role": "assistant", "content": "Is there any discharge?"}, {"role": "user", "content": "Yes"}, {"role": "user", "content": "Please check my description, I wrote all the details"}, {"role": "assistant", "content": "Sure"}, {"role": "assistant", "content": "The internet was down just now"}, {"role": "user", "content": "Okay"}, {"role": "assistant", "content": "Is the discharge thick, thin, or stringy?"}, {"role": "user", "content": "It's thick"}, {"role": "user", "content": "Yellowish"}, {"role": "user", "content": "Mainly a lot in the morning, and a lot of phlegm"}, {"role": "assistant", "content": "Does it affect your vision? Do you have eye pain? Itchy eyes? Foreign body sensation? Tears?"}, {"role": "user", "content": "No"}, {"role": "user", "content": "Mainly still sore throat"}, {"role": "user", "content": "The eyes are just red and have discharge"}, {"role": "user", "content": "Sore throat, a lot of phlegm, mild cough, hoarse voice"}, {"role": "assistant", "content": "Okay"}, {"role": "assistant", "content": "Have you had any medical examinations or medication history? Any history of drug allergies or chronic diseases?"}, {"role": "user", "content": "No"}, {"role": "user", "content": "Please help as soon as possible, it's getting late"}, {"role": "assistant", "content": "<med_search>"}]
+}
+```
+### Extract ground-truth labels for rewards
+From `remaining_chat`, derive the following new fields:
+- `decision_truth`: the correct action (e.g., `"continue"` or `"stop"`),
+- `info_truth`: structured symptom information used for reward computation
+These ground truth are used to evaluate the rewards in training, e.g., $R_a$ and $R_s$.
+```JSON
+{
+   "cid": "35310_7",
+   "session_id": "35310",
+   "diagn": "Upper Respiratory Tract Infection",
+   "messages": [{"role": "user", "content": "Sore throat, phlegm, red eyes, cough, hoarse voice"}, {"role": "user", "content": "I took Amoxicillin"}, {"role": "user", "content": "But I still don't feel well"}, {"role": "user", "content": "Mainly it's a respiratory infection, sore throat, phlegm, hoarse voice, red eyes"}, {"role": "user", "content": "When I wake up, there is a lot of eye discharge, and a lot of phlegm"}, {"role": "assistant", "content": "How long have the symptoms been present?"}, {"role": "user", "content": "About 2 days"}, {"role": "user", "content": "My eyes are very red"}],
+   "remaining_chat": [{"role": "assistant", "content": "Is there any discharge?"}, {"role": "user", "content": "Yes"}, {"role": "user", "content": "Please check my description, I wrote all the details"}, {"role": "assistant", "content": "Sure"}, {"role": "assistant", "content": "The internet was down just now"}, {"role": "user", "content": "Okay"}, {"role": "assistant", "content": "Is the discharge thick, thin, or stringy?"}, {"role": "user", "content": "It's thick"}, {"role": "user", "content": "Yellowish"}, {"role": "user", "content": "Mainly a lot in the morning, and a lot of phlegm"}, {"role": "assistant", "content": "Does it affect your vision? Do you have eye pain? Itchy eyes? Foreign body sensation? Tears?"}, {"role": "user", "content": "No"}, {"role": "user", "content": "Mainly still sore throat"}, {"role": "user", "content": "The eyes are just red and have discharge"}, {"role": "user", "content": "Sore throat, a lot of phlegm, mild cough, hoarse voice"}, {"role": "assistant", "content": "Okay"}, {"role": "assistant", "content": "Have you had any medical examinations or medication history? Any history of drug allergies or chronic diseases?"}, {"role": "user", "content": "No"}, {"role": "user", "content": "Please help as soon as possible, it's getting late"}, {"role": "assistant", "content": "<med_search>"}],
+   "decision_truth": "continue",
+   "info_truth": "Symptom: sore throat, Symptom quality: thick discharge, Symptom quality: yellowish discharge, Symptom quality: a lot of phlegm, Symptom severity: mild cough, Symptom quality: hoarse voice"
+}
+```
+
+---
+
+## Step 2. Configure and Train
+Update `examples/learn_to_ask/train.yaml` with paths to:
+- Your processed datasets,
+- Base model,
+- Checkpoint output directory.
+
+Here is an example configuration:
+```yaml
+mode: both
+project: learn2ask
+name: learn2ask_example
+checkpoint_root_dir: ${oc.env:TRINITY_CHECKPOINT_ROOT_DIR,./checkpoints}  # Checkpoint output directory
+# some configs...
+model:
+  model_path: ${oc.env:TRINITY_MODEL_PATH,Qwen/Qwen2.5-7B-Instruct}  # Base model
+# some configs...
+buffer:
+  batch_size: 64
+  total_epochs: 4
+  explorer_input:
+    taskset:
+      name: taskset
+      storage_type: file
+      path: ${oc.env:TRINITY_TASKSET_PATH,examples/learn_to_ask/data}  # Your processed datasets
+      split: train
+      subset_name: null
+      format:
+        prompt_key: messages
+        response_key: action_truth
+      workflow_args:  # Workflow arguments
+        train_mode: "Ra+Rs"
+        fusion_mode: "default"
+# some configs...
+explorer:
+  # some configs...
+  auxiliary_models:
+    - model_path: ${oc.env:TRINITY_AUX_MODEL_PATH,Qwen/Qwen2.5-32B-Instruct}  # Auxiliary model path
+# some configs...
+```
+
+Then, launch training:
+```bash
+trinity run --config examples/learn_to_ask/train.yaml --plugin-dir examples/learn_to_ask/workflow
+```
+---
+
+## Step 3. Evaluate
+Use the rollout-n-evaluate pipeline:
+- Generate responses on the test set (e.g. with *vLLM*),
+- Evaluate outputs using **`qwen2.5-32b-instruct`** with the evaluator.
+
+You may configure the settings then run the pipeline by launching:
+```bash
+python examples/learn_to_ask/data_prepare/3_rollout_then_evaluate.py --eval_model_path path/to/trained/model --grader_model_path path/to/qwen2.5-32b-instruct  --test_file_path examples/learn_to_ask/data/test.jsonl --rollout_file_path path/to/rollout.jsonl --eval_file_path path/to/output.jsonl
+```
+
+Note: `eval_model_path` is the location of the model you want to evaluate. This model must first be converted into the HuggingFace format. For instructions on converting FSDP checkpoints, see [this guide](https://modelscope.github.io/Trinity-RFT/en/main/tutorial/faq.html).
+
+
+## Citation
+
+If you find this code useful, please consider citing our papers:
+
+```bibtex
+@misc{learn2ask,
+      title={Grounded in Reality: Learning and Deploying Proactive LLM from Offline Logs},
+      author={Fei Wei and Daoyuan Chen and Ce Wang and Yilun Huang and Yushuo Chen and Xuchen Pan and Yaliang Li and Bolin Ding},
+      year={2025},
+      eprint={2510.25441},
+      archievePrefix={arXiv},
+      primaryClass={cs.AI},
+      url={https://arxiv.org/abs/2510.25441}
+}
+```

examples/learn_to_ask/data_prepare/1_info_extract_pipeline.py

@@ -0,0 +1,132 @@
+import argparse
+import json
+import time
+
+from llm_info_extraction import LLM_info_extraction, parse_llm_output
+from message_splitter import split_session_to_json_lines
+
+
+def process_jsonl_file(
+    input_file, output_file, model_call_mode="online_api", max_retries=3, **kwargs
+):
+    """
+    Process all sessions in a JSONL file and save results to output file.
+
+    Args:
+        input_file (str): Path to input JSONL file
+        output_file (str): Path to output JSONL file
+        model_call_mode (str): Either "online_api" or "local_vllm"
+        max_retries (int): Maximum number of retries for LLM calls
+        **kwargs: Additional parameters for API calls
+
+    Returns:
+        str: Success message or error information
+    """
+    try:
+        # Read and process each session
+        with open(input_file, "r", encoding="utf-8") as infile, open(
+            output_file, "w", encoding="utf-8"
+        ) as outfile:
+            for line_num, line in enumerate(infile, 1):
+                if line.strip():
+                    try:
+                        session = json.loads(line)
+                        print(
+                            f"Processing session {session.get('session_id', 'unknown')} (line {line_num})..."
+                        )
+
+                        # Process the session
+                        processed_lines = process_session(
+                            session, model_call_mode, max_retries, **kwargs
+                        )
+                        for processed_line in processed_lines:
+                            outfile.write(processed_line + "\n")
+
+                    except json.JSONDecodeError as e:
+                        print(f"Warning: Skipping invalid JSON at line {line_num}: {e}")
+                    except Exception as e:
+                        print(f"Warning: Error processing session at line {line_num}: {e}")
+
+        return f"Successfully processed. Results saved to {output_file}"
+
+    except Exception as e:
+        return f"Error processing JSONL file: {str(e)}"
+
+
+def process_session(session, model_call_mode="online_api", max_retries=3, **kwargs):
+    """
+    Pipeline function that splits messages into rounds and extracts info from each round's remaining chat.
+
+    Args:
+        session (dict): Session dictionary containing 'session_id', 'diagn', and 'messages' keys
+        model_call_mode (str): Either "online_api" or "local_vllm"
+        max_retries (int): Maximum number of retries for LLM calls
+        **kwargs: Additional parameters for API calls
+
+    Returns:
+        list: List of JSON strings with added "info_set" key, or error information
+    """
+    try:
+        # Step 1: Split messages into JSON lines
+        json_lines = split_session_to_json_lines(session)
+
+        # Step 2: Process each JSON line with LLM info extraction
+        processed_lines = []
+
+        for line in json_lines:
+            data = json.loads(line)
+            remaining_chat = data.get("remaining_chat", "")
+
+            # Retry loop for LLM calls
+            info_set = None
+            for attempt in range(max_retries):
+                try:
+                    # Call LLM info extraction (using mock function for testing)
+                    llm_response = LLM_info_extraction(remaining_chat, model_call_mode, **kwargs)
+
+                    info_set = parse_llm_output(llm_response)
+
+                    if isinstance(info_set, list):
+                        break
+                    else:
+                        # If parsing failed, this is an error message
+                        print(f"Attempt {attempt + 1} failed: {info_set}")
+                        if attempt < max_retries - 1:
+                            time.sleep(1)
+                except Exception as e:
+                    print(f"Attempt {attempt + 1} failed with exception: {str(e)}")
+                    if attempt < max_retries - 1:
+                        time.sleep(1)  # Shorter wait for testing
+
+            data["info_set"] = info_set
+            processed_lines.append(json.dumps(data, ensure_ascii=False))
+
+        return processed_lines
+
+    except Exception as e:
+        return f"Pipeline error: {str(e)}"
+
+
+# Example usage:
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--input_file", type=str, default="examples/learn_to_ask/data_raw/train_origin.jsonl"
+    )
+    parser.add_argument(
+        "--output_file", type=str, default="examples/learn_to_ask/data_raw/train_processed.jsonl"
+    )
+    parser.add_argument(
+        "--model_call_mode", type=str, choices=["online_api", "local_vllm"], default="local_vllm"
+    )
+    parser.add_argument("--model_path", type=str, required=True)
+    args = parser.parse_args()
+    print(
+        process_jsonl_file(
+            input_file=args.input_file,
+            output_file=args.output_file,
+            model_call_mode=args.model_call_mode,
+            model_path=args.model_path,
+            # Additional parameters for API calls
+        )
+    )

examples/learn_to_ask/data_prepare/2_build_dataset.py

@@ -0,0 +1,55 @@
+import argparse
+import json
+
+
+def process_message(json_obj):
+    info_set = json_obj.get("info_set")
+    info_set_str = ", ".join(info_set) if isinstance(info_set, list) else ""
+    if "user: " not in json_obj["remaining_chat"]:
+        decision_str = "stop"
+    else:
+        decision_str = "continue"
+    if not info_set_str and decision_str == "continue":
+        if_keep = False
+    else:
+        if_keep = True
+    return if_keep, info_set_str, decision_str
+
+
+def main(input_file_path, output_file_path):
+    with open(input_file_path, "r", encoding="utf-8") as infile, open(
+        output_file_path, "w", encoding="utf-8"
+    ) as outfile:
+        print("data processing started...")
+        for line in infile:
+            data = json.loads(line.strip())
+            if_keep, info_set, decision = process_message(data)
+            if not if_keep:
+                continue
+
+            new_item = {
+                "cid": data["cid"],
+                "session_id": data["session_id"],
+                "diagn": data["diagn"],
+                "messages": data["messages"],
+                "decision_truth": decision,
+                "info_truth": info_set,
+            }
+            outfile.write(json.dumps(new_item, ensure_ascii=False) + "\n")
+    print("job done!")
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+
+    # The file generated by 1_info_extract_pipeline.py
+    parser.add_argument(
+        "--input_file", type=str, default="examples/learn_to_ask/data_raw/train_processed.jsonl"
+    )
+
+    # The final file for training or testing
+    parser.add_argument("--output_file", type=str, default="examples/learn_to_ask/data/train.jsonl")
+
+    args = parser.parse_args()
+
+    main(args.input_file, args.output_file)