File size: 12,678 Bytes
b4806b0
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
"""
EHRGym GRPO Reinforcement Learning Script  (Unsloth)
=====================================================
Trains a model with GRPO to interact with the live EHRGym environment.
The model learns to navigate the EHR, place orders, and write notes
by receiving reward signals from the rubric evaluator.

Prerequisites:
  1. EHRGym running:  npm run dev          (Next.js + env server)
  2. Dependencies:    pip install -r requirements-train.txt

Usage (on H100):
  python scripts/train_grpo.py \
      --model unsloth/Qwen2.5-7B-Instruct \
      --output runs/checkpoints/ehrgym-grpo \
      --max-steps 500 --num-generations 4

Quick smoke test:
  python scripts/train_grpo.py \
      --model unsloth/Qwen2.5-0.5B-Instruct \
      --output runs/checkpoints/ehrgym-grpo-tiny \
      --max-steps 20 --num-generations 2 --lora-r 16
"""
from __future__ import annotations

import argparse
import json
import logging
import os
from pathlib import Path

logging.basicConfig(level=logging.INFO, format="%(asctime)s %(levelname)s %(message)s")
log = logging.getLogger(__name__)

# ---------------------------------------------------------------------------
# Constants
# ---------------------------------------------------------------------------

SYSTEM_PROMPT = (
    "You are a clinical computer-use agent operating an Epic-like EHR. "
    "Given the current screenshot description, URL, activity, task goal, and state metadata, "
    "return exactly one valid next action as strict JSON.\n\n"
    "Valid action types: click, fill, keypress, goto, wait.\n"
    "Examples:\n"
    '  {"type": "click", "selector": "[data-testid=\'order-btn\']"}\n'
    '  {"type": "fill", "selector": "#note-body", "value": "Patient improving..."}\n'
    '  {"type": "goto", "url": "http://127.0.0.1:3000/patient/pat-1001"}\n'
)

ENV_SERVER = "http://127.0.0.1:8000"
TASK_ID = "aki-chart-review"  # default task

# ---------------------------------------------------------------------------
# CLI
# ---------------------------------------------------------------------------

def parse_args() -> argparse.Namespace:
    p = argparse.ArgumentParser(description="EHRGym GRPO RL training (Unsloth)")
    p.add_argument("--model", default="unsloth/Qwen2.5-7B-Instruct")
    p.add_argument("--output", default="runs/checkpoints/ehrgym-grpo")
    p.add_argument("--max-steps", type=int, default=500)
    p.add_argument("--num-generations", type=int, default=4,
                   help="Number of completions to sample per prompt (GRPO group size)")
    p.add_argument("--max-seq-len", type=int, default=2048)
    p.add_argument("--lora-r", type=int, default=64)
    p.add_argument("--lora-alpha", type=int, default=128)
    p.add_argument("--lr", type=float, default=5e-5)
    p.add_argument("--batch-size", type=int, default=1)
    p.add_argument("--grad-accum", type=int, default=4)
    p.add_argument("--no-4bit", action="store_true")
    p.add_argument("--env-server", default=ENV_SERVER)
    p.add_argument("--task-id", default=TASK_ID)
    p.add_argument("--max-episode-steps", type=int, default=25,
                   help="Max env steps per episode before termination")
    p.add_argument("--seed", type=int, default=3407)
    p.add_argument("--wandb-project", default=None)
    p.add_argument("--save-method", default="lora",
                   choices=["lora", "merged_16bit", "merged_4bit"])
    return p.parse_args()


# ---------------------------------------------------------------------------
# Environment interaction helpers
# ---------------------------------------------------------------------------

def env_reset(base_url: str, task_id: str) -> dict:
    """Reset the EHRGym environment and return the initial observation."""
    import httpx
    resp = httpx.post(f"{base_url}/reset", json={"task_id": task_id}, timeout=30)
    resp.raise_for_status()
    return resp.json()


def env_step(base_url: str, action: dict) -> dict:
    """Take an action in the EHRGym environment."""
    import httpx
    resp = httpx.post(f"{base_url}/step", json=action, timeout=30)
    resp.raise_for_status()
    return resp.json()


def obs_to_text(obs: dict) -> str:
    """Convert an EHRGym observation to a text prompt (no screenshot b64)."""
    payload = {
        "goal": obs.get("goal", ""),
        "current_url": obs.get("current_url", ""),
        "active_activity": obs.get("active_activity", ""),
        "state": obs.get("state", {}),
    }
    return json.dumps(payload, ensure_ascii=False)


# ---------------------------------------------------------------------------
# Reward functions
# ---------------------------------------------------------------------------

def valid_json_reward(completions: list, **kwargs) -> list[float]:
    """Reward: is the model output valid JSON with a 'type' field?

    Scale: [-1.0, +0.5]  (format correctness is a prerequisite, not the goal)
    """
    scores = []
    for completion in completions:
        text = completion[0]["content"] if isinstance(completion, list) else completion
        try:
            parsed = json.loads(text)
            if "type" in parsed:
                scores.append(0.5)
            else:
                scores.append(-0.5)
        except (json.JSONDecodeError, TypeError):
            scores.append(-1.0)
    return scores


def action_type_reward(completions: list, **kwargs) -> list[float]:
    """Reward: does the action use a valid type AND include required fields?

    Scale: [-0.5, +0.5]  (valid action structure is necessary but not sufficient)
    """
    valid_types = {"click", "fill", "keypress", "goto", "wait"}
    required_fields = {
        "click": ["selector"], "fill": ["selector", "text"],
        "keypress": ["key"], "goto": ["url"], "wait": [],
    }
    scores = []
    for completion in completions:
        text = completion[0]["content"] if isinstance(completion, list) else completion
        try:
            parsed = json.loads(text)
            action_type = parsed.get("type")
            if action_type not in valid_types:
                scores.append(-0.5)
                continue
            # Check that required fields are present and non-empty
            fields = required_fields.get(action_type, [])
            if all(parsed.get(f) for f in fields):
                scores.append(0.5)
            else:
                scores.append(0.0)  # right type but incomplete
        except Exception:
            scores.append(-0.5)
    return scores


def rubric_progress_reward(completions: list, **kwargs) -> list[float]:
    """
    Reward: execute the action against the live env and return rubric reward.
    This is the main task reward — it actually steps the environment.
    
    For each completion we reset the env first so every candidate starts from
    the same state (critical for GRPO where multiple completions share the same
    prompt).  The env now returns an incremental reward and a breakdown in info.
    """
    env_url = kwargs.get("env_url", ENV_SERVER)
    task_id = kwargs.get("task_id", TASK_ID)
    
    scores = []
    for completion in completions:
        text = completion[0]["content"] if isinstance(completion, list) else completion
        try:
            action = json.loads(text)
        except Exception:
            scores.append(-2.0)
            continue
        
        try:
            # Reset before each completion so state is clean
            env_reset(env_url, task_id)
            # Step the environment with this single action
            result = env_step(env_url, action)
            # The env now returns a well-calibrated incremental reward
            reward = result.get("reward", 0.0)
            # Amplify to make the rubric signal dominant over format rewards
            scores.append(reward * 5.0)
        except Exception as e:
            log.warning("Env step failed: %s", e)
            scores.append(-1.0)
    
    return scores


# ---------------------------------------------------------------------------
# Dataset: generate prompts by resetting env
# ---------------------------------------------------------------------------

def build_prompt_dataset(env_url: str, task_id: str, n_prompts: int = 100):
    """Create a dataset of prompts by resetting the env multiple times."""
    from datasets import Dataset
    
    rows = []
    for i in range(n_prompts):
        try:
            obs = env_reset(env_url, task_id)
            observation = obs.get("observation", obs)
            user_content = obs_to_text(observation)
            rows.append({
                "prompt": [
                    {"role": "system", "content": SYSTEM_PROMPT},
                    {"role": "user", "content": user_content},
                ],
            })
        except Exception as e:
            log.warning("Failed to reset env (prompt %d): %s", i, e)
            continue
    
    if not rows:
        raise RuntimeError(f"Could not get any prompts from {env_url}")
    
    log.info("Built %d prompts from env resets", len(rows))
    return Dataset.from_list(rows)


# ---------------------------------------------------------------------------
# Main
# ---------------------------------------------------------------------------

def main() -> None:
    args = parse_args()

    from unsloth import FastLanguageModel
    from unsloth.chat_templates import get_chat_template
    from trl import GRPOConfig, GRPOTrainer

    # ---- W&B ----
    if args.wandb_project:
        os.environ["WANDB_PROJECT"] = args.wandb_project
        report_to = "wandb"
    else:
        report_to = "none"

    # ---- Load model ----
    load_in_4bit = not args.no_4bit
    model, tokenizer = FastLanguageModel.from_pretrained(
        model_name=args.model,
        max_seq_length=args.max_seq_len,
        load_in_4bit=load_in_4bit,
        dtype=None,
    )

    model = FastLanguageModel.get_peft_model(
        model,
        r=args.lora_r,
        lora_alpha=args.lora_alpha,
        lora_dropout=0.0,
        target_modules=[
            "q_proj", "k_proj", "v_proj", "o_proj",
            "gate_proj", "up_proj", "down_proj",
        ],
        use_gradient_checkpointing="unsloth",
        random_state=args.seed,
    )
    model.print_trainable_parameters()

    # Chat template
    model_lower = args.model.lower()
    if "qwen" in model_lower:
        chat_template = "qwen-2.5"
    elif "llama" in model_lower:
        chat_template = "llama-3.1"
    else:
        chat_template = "chatml"
    tokenizer = get_chat_template(tokenizer, chat_template=chat_template)

    # ---- Dataset (prompts from env) ----
    dataset = build_prompt_dataset(args.env_server, args.task_id, n_prompts=200)

    # ---- GRPO config ----
    max_prompt_length = args.max_seq_len // 2
    max_completion_length = args.max_seq_len - max_prompt_length

    output_dir = Path(args.output)
    output_dir.mkdir(parents=True, exist_ok=True)

    training_args = GRPOConfig(
        output_dir=str(output_dir),
        max_steps=args.max_steps,
        per_device_train_batch_size=args.batch_size,
        gradient_accumulation_steps=args.grad_accum,
        num_generations=args.num_generations,
        max_prompt_length=max_prompt_length,
        max_completion_length=max_completion_length,
        learning_rate=args.lr,
        lr_scheduler_type="linear",
        warmup_ratio=0.1,
        weight_decay=0.01,
        optim="adamw_8bit",
        bf16=True,
        logging_steps=1,
        save_steps=100,
        save_total_limit=3,
        seed=args.seed,
        report_to=report_to,
        temperature=1.0,
    )

    trainer = GRPOTrainer(
        model=model,
        processing_class=tokenizer,
        args=training_args,
        train_dataset=dataset,
        reward_funcs=[
            valid_json_reward,       # [-1.0, +0.5] format correctness
            action_type_reward,      # [-0.5, +0.5] valid type + required fields
            rubric_progress_reward,  # [-2.0, +5.0] actual task rubric (dominant)
        ],
    )

    # ---- Train ----
    log.info("Starting GRPO training (max_steps=%d, num_generations=%d) …",
             args.max_steps, args.num_generations)
    trainer.train()

    # ---- Save ----
    if args.save_method == "lora":
        save_dir = str(output_dir / "lora_adapter")
        model.save_pretrained(save_dir)
        tokenizer.save_pretrained(save_dir)
        log.info("Saved LoRA adapter → %s", save_dir)
    else:
        save_dir = str(output_dir / "merged")
        model.save_pretrained_merged(save_dir, tokenizer, save_method=args.save_method)
        log.info("Saved merged model → %s", save_dir)

    log.info("Done ✓")


if __name__ == "__main__":
    main()