GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
RP-360
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
RP-360
Model Tree
RP-360/Qwen3.5-4B-mdr-narrative-sft-codes-weighted
Input Modalities
Output Modalities
Supported Functionality
RP-360/Qwen3.5-4B-mdr-narrative-grpo-codes-weighted represents the reinforcement learning alignment stage of the Medical Device Report (MDR) Narrative Intelligence system. Trained directly on top of RP-360/Qwen3.5-4B-mdr-narrative-sft-codes-weighted, this policy model optimizes structured clinical information extraction through 300 steps of Group Relative Policy Optimization (GRPO) (2,400 rollouts, group size G = 8) guided by a 7-Dimensional Programmatic Reward Suite.
0.4538 vs Constant Baseline 0.1753 (+158.8% relative gain, Precision: 48.63%, Recall: 44.29%, Exact Match: 39.50%).0.66780.260069.27%66.62%56.50%0.6678)0.5608 (+157.6% relative gain over baseline 0.2177), outperforming the pre-registered GRPO ROI hurdle rate of 0.3383 by +65.8%.99.56% citation precision (448/450 quotes verbatim verified against source narratives; 0.44% hallucination rate), guarded by hard invariant penalties.mni-json-v2-codes with zero missing keys.+0.6344 across held-out reports (98.0% positive reward steps, 37.0% high quality ≥ +0.70).
Evaluation was performed on a standalone benchmark of 200 held-out medical adverse event reports from RP-360/mdr-narrative-consolidated-pre-20k. A strict data partition audit verified 0 overlap with the training corpus (zero data leakage).
| Evaluation Dimension | Constant Baseline | GRPO Hurdle Target | Qwen3.5-4B GRPO Policy | Absolute Delta | Relative Gain | Clinical Status |
|---|---|---|---|---|---|---|
| Annex A Macro F1 (Device Problems) | 0.1753 | — | 0.4538 | +0.2785 | +158.8% | 🟢 Outperforms Baseline |
| Annex E Macro F1 (Patient Harms) | 0.2600 | — | 0.6678 | +0.4078 | +156.9% | 🟢 Reaches Head Data Ceiling |
| Combined Macro F1 (Overall IMDRF) | 0.2177 | 0.3383 | 0.5608 | +0.3431 | +157.6% | 🎯 Beats Hurdle (+65.8%) |
| Schema Pass Rate | 0.0% | 95.0% | 98.5% (197/200) | — | — | 🟢 Strict JSON Compliance |
| Verbatim Quote Grounding | 0.0% | 90.0% | 99.56% (448/450) | — | — | 🛡️ 0.44% Hallucination Rate |
Categorical: event_type | 53.5% | — | 90.0% (180/200) | +36.5% | — | 🟢 High-Confidence Class. |
Categorical: harm_bucket | 76.0% | — | 90.5% (181/200) | +14.5% | — | 🟢 High Sensitivity to Harm |
Categorical: DEVICE_EXAMINED | 61.5% | — | 72.0% (144/200) | +10.5% | — | 🟢 Consistent Adjudication |
| Mean Composite Policy Reward | 0.0000 | > +0.3000 | +0.6344 | +0.6344 | — | 🚀 Optimal Policy Convergence |
The model predicts three critical categorical metadata attributes: DEVICE_EXAMINED (whether the manufacturer physical lab inspected the device), event_type (Malfunction vs Injury vs Death), and harm_bucket (harm vs no_harm).
Each predicted IMDRF code requires supporting narrative evidence formatted as {"code": "...", "quote": "..."}. During evaluation, all emitted quotes were tested for exact substring presence in the source reporter and manufacturer narratives:
| Evidence Grounding Dimension | Metric Value | Benchmark Standard | Status |
|---|---|---|---|
| Total Emitted Evidence Quotes | 450 | — | Emitted across 200 reports |
| Verbatim Grounded Quotes | 448 | ≥ 405 (90.0%) | 🟢 Substring-level match in raw report |
| Ungrounded Quotes (Hallucinations) | 2 | ≤ 45 (10.0%) | 🛡️ 0.44% Hallucination Rate |
| Grounding Precision Rate | 99.56% | ≥ 90.0% | 🛡️ Enforced by hard -0.20 INV8 penalty |
| Mean Quote Token Length | 10.05 words | 5 – 25 words | 🟢 Concise, semantically focused quotes |
The table below compares the empirical baseline, both Supervised Fine-Tuning (SFT) generations, and the final GRPO reinforcement learning policy across all 200 held-out validation reports:
| Evaluation Dimension | Constant / Majority Baseline | Gen 1 SFT Model: Full JSON (-sft-json) | Gen 2 SFT Model: Codes-Weighted (-sft-codes-weighted) | Gen 3 GRPO RL Policy (-grpo-codes-weighted) | Empirical Ceiling / Target | Advantage & RL Impact |
|---|---|---|---|---|---|---|
| Schema Syntax Version | — | mni-json-v1 (embedded defs) | mni-json-v2-codes (codes only) | mni-json-v2-codes | — | Optimized token budget (~390 tokens/report) |
| Generation Latency (200 reports) | — | 1,127s (18m 46s) | 618s (10m 18s) | 497s (8m 17s) | — | ⚡ 1.24x Faster vs Gen 2, 2.27x vs Gen 1 |
| Throughput (reports/sec) | — | 0.177 | 0.324 | 0.402 | — | ⚡ +24.1% higher generation throughput |
| Schema Pass Rate | 0.0% | 99.0% (198/200) | 99.0% (198/200) | 98.5% (197/200) | 100.0% | 🟢 Flawless syntactic reliability |
| Annex A Macro F1 (Device Problems) | 0.1753 | 0.4180 (69.2% ceiling) | 0.4360 (72.2% ceiling) | 0.4538 (76.3% ceiling) | 0.5947 | 🟢 Highest Annex A F1 across all models |
| Annex E Macro F1 (Patient Harms) | 0.2600 | 0.6360 (95.2% ceiling) | 0.6250 (93.6% ceiling) | 0.6678 (100.0% ceiling) | 0.6678 | 🟢 Reaches Empirical Head Ceiling |
| Combined Macro F1 (Clinical Overall) | 0.2177 | 0.5270 | 0.5305 | 0.5608 | — | 🎯 Beats GRPO Hurdle (0.3383) by +65.8% |
| Average Composite Policy Reward | 0.0000 | N/A (unaligned) | -0.4370 (initial rollout) | +0.6344 (Peak: +0.9941) | +1.0000 | 🚀 Monotonic RL trajectory convergence |
| Taxonomy Terms + Defs Fidelity | 0.0% | 88.7% (param memory) | 99.6% (hydration) | 99.6% (hydration) | 100.0% | 🟢 Zero Definition Hallucination |
| Evidence Quotes Grounding Precision | 0.0% | 99.3% (398/401 verbatim) | 99.3% (445/448 verbatim) | 99.56% (448/450 verbatim) | 100.0% | 🛡️ Guarded by hard -0.20 INV8 penalty |
| Adverse Event Type Accuracy | 53.5% | 90.5% | 89.0% | 90.0% | 100.0% | 🟢 +36.5% over majority class |
| Harm Severity Bucket Accuracy | 76.0% | 89.5% | 90.0% | 90.5% | 100.0% | 🟢 +14.5% over majority class |
| Device Examination Accuracy | 61.5% | 64.0% | 62.5% | 72.0% | 100.0% | 🟢 +9.5% gain over SFT (Highest) |
| Anti-Gaming Guardrails | None | None | None | Active (AH, INV1, INV8, INV9) | Required | 🛡️ Prevents code flooding & negations |
The GRPO policy model is built upon a two-stage Supervised Fine-Tuning (SFT) foundation prior to reinforcement learning:
-sft-json)The foundational clinical extraction model (RP-360/Qwen3.5-4B-mdr-narrative-sft-json) was trained on 9,087 reports on an NVIDIA A100-SXM4 GPU (2.54 h) with early stopping triggered at Step 1,600 / 2,835 (Epoch 2.82). Checkpoint 1100 reached the global minimum validation loss of 0.12582 across all 431 held-out validation reports:
-sft-codes-weighted)Warm-starting directly from checkpoint-1100, the model was adapted to the streamlined mni-json-v2-codes syntax with differential token loss weights:
DEVICE_EXAMINED, event_type, harm_bucket).think, FAILURE_MODE, MECHANISM, PATIENT_HARM, etc.).
Trained with effective batch size 16 (sequence length 3,072) on an NVIDIA A100-SXM4 GPU with early stopping (patience 5). Evaluated every 100 steps across all 431 held-out validation reports:
| Evaluation Step | Training Loss (Cross-Entropy) | Validation Loss (431 Reports) | Delta vs Prior | Status / Checkpoint |
|---|---|---|---|---|
| Step 100 | 0.8535 | 0.24775 | — | Initial validation evaluation |
| Step 200 | 0.7250 | 0.24981 | +0.00206 | Patience 1/5 |
| Step 300 | 0.7358 | 0.24451 | -0.00530 | 🟢 Improvement |
| Step 400 | 0.7329 | 0.24863 | +0.00412 | Patience 1/5 |
| Step 500 | 0.7208 | 0.24415 | -0.00448 | 🏆 Optimal Checkpoint (Merged as Base Weights for GRPO) |
| Step 600 | 0.6384 | 0.25513 | +0.01098 | Patience 1/5 (Validation loss plateauing) |
| Step 700 | 0.5849 | 0.26128 | +0.00615 | Patience 2/5 |
| Step 800 | 0.5848 | 0.26515 | +0.00387 | Patience 3/5 |
| Step 900 | 0.6587 | 0.26179 | -0.00336 | Patience 4/5 |
| Step 1,000 | 0.5878 | 0.26232 | +0.00053 | ⏹️ Early Stopping Triggered (Patience 5/5 reached) |
checkpoint-500 achieved the lowest validation cross-entropy loss (0.24415) and was selected and merged as the starting weights for GRPO policy alignment.
Before launching GRPO, the base policy RP-360/Qwen3.5-4B-mdr-narrative-sft-codes-weighted underwent multi-temperature latent exploration diagnostics (Pass@K) to establish the empirical ROI hurdle rate. The table below traces the diagnostic benchmarks to the final post-GRPO policy results:
| Diagnostic Dimension / Check | Pre-GRPO Value | Benchmark Meaning | Post-GRPO Actual Policy | Status & Achievement |
|---|---|---|---|---|
| Pass@1 (F1 ≥ 0.50) | 0.135 | Single-Pass Greedy Floor (SFT base) | 0.5608 | 🟢 +0.426 jump over greedy base |
| Pass@8 (F1 ≥ 0.50) | 0.480 | Upper Bound Exploration Spread | 0.5608 | 🚀 Surpasses Pass@8 latent ceiling (+0.081) |
| Exploration Variance (+Δ) | +0.345 | Spread required for RL (≥ +0.18) | — | 🟢 Sufficient exploratory headroom verified |
| Self-Consistency 8-Vote Baseline | 0.3383 | GRPO ROI Hurdle Rate to Beat | 0.5608 | 🎯 Beats ROI Hurdle Rate by +65.8% |
| Pass@8 Exact Full-Record Match | 0.280 | Peak Exact Match Exploration | 0.395 (Annex A) / 0.565 (Annex E) | 🟢 Significant leap in exact match recall |
| Curated Prime RL Pool | 12 (48.0%) | Goldilocks Active Learning candidates | 300 steps | 🟢 Optimized across hard clinical cases |
| Peak Single-Rollout Reward | -0.437 | Cold-start baseline step 1 | +0.9941 | 🏆 Near-flawless convergence at Step 286 |
Rather than relying on opaque neural reward models that are susceptible to reward gaming, this GRPO alignment uses an authoritative, rule-based clinical verifier evaluating 7 orthogonal dimensions:
text
┌────────────────────────────────────────────────────────────────────────┐│ Composite Reward Function │├────────────────────────────────────────────────────────────────────────┤│ R = clip( ∑ (w_k · C_k) - Total_Penalties, min = -1.0, max = +1.0 )│└────────────────────────────────────────────────────────────────────────┘text
┌────────────────────────────────────────────────────────────────────────┐│ 7-Dimensional Clinical Reward Suite │├────────────────────────────────────────────────────────────────────────┤│ 1. IMDRF Annex A F1 Score (Device Problems) [22%] ││ 2. IMDRF Annex E F1 Score (Health Effects / Patient Harm) [22%] ││ 3. Annex Harmonic Balance (min(A, E)) [10%] ││ 4. Grounded Citation Verbatim Match Precision [12%] ││ 5. Citation Length & Semantic Quality [06%] ││ 6. Categorical Consistency (Harm Bucket, Event Type, Exam) [20%] ││ 7. Causal Chain Reasoning (Failure -> Mechanism -> Outcome) [08%] │└────────────────────────────────────────────────────────────────────────┘To prevent the policy from exploiting reward formulation shortcuts (such as code spamming or citing non-existent sentences), 5 strict invariants are enforced during rollout scoring:
EVIDENCE is not found verbatim in the source report, an immediate penalty of -0.20 is applied.1069), an immediate invariant violation penalty is triggered.harm_bucket == "harm" but no Annex E code is provided, a -0.20 penalty is levied.GRPO optimization demonstrated monotonic convergence across all 300 steps on an NVIDIA A100-SXM4 GPU.
| Step Window | Learning Rate | Avg Reward | % Positive Steps | High Quality (Reward ≥ +0.70) | Avg KL Divergence | Convergence Phase |
|---|---|---|---|---|---|---|
| Steps 1 – 20 | 0 → 3.0 × 10⁻⁶ | -0.437 | 10.0% | 0.0% | 0.0017 | Cold-start / exploration baseline |
| Steps 21 – 40 | 3.0 → 4.8 × 10⁻⁶ | +0.034 | 60.0% | 0.0% | 0.0873 | Inflection point: Policy shifts positive |
| Steps 41 – 60 | Peak 5.0 × 10⁻⁶ | +0.207 | 65.0% | 0.0% | 0.1514 | Syntax & schema stabilization |
| Steps 61 – 80 | 4.8 → 4.6 × 10⁻⁶ | +0.431 | 90.0% | 20.0% | 0.1890 | Clinical code precision calibration |
| Steps 81 – 100 | 4.6 → 4.2 × 10⁻⁶ | +0.681 | 100.0% | 60.0% | 0.2186 | Active clinical mastery |
| Steps 151 – 176 | 2.9 → 2.2 × 10⁻⁶ | +0.687 | 96.2% | 65.4% | 0.2257 | Sustained precision across rare devices |
| Steps 280 – 300 | Annealed → ~0 | +0.742 | 100.0% | 70.0% | 0.2584 | Peak convergence: Hits +0.9941 |
0.0005 during initial warmup to settle stably between 0.20 and 0.26.
The training and evaluation runs executed on an NVIDIA A100-SXM4-40GB:
mni-json-v2-codes)The model emits deterministic JSON where IMDRF codes are produced purely as alphanumeric strings, saving ~87 tokens per generation:
json
{ "think": "Patient presented with lead migration resulting in loss of pacing capture...", "FAILURE_MODE": "Lead displacement", "MECHANISM": "Electrode dislodgement -> loss of capture", "PATIENT_HARM": "Syncope and recurrent arrhythmia", "INTERVENTION": "Surgical revision and lead replacement", "OUTCOME": "Recovered post-reoperation", "DEVICE_ROLE": "Directly responsible", "IMDRF_A": ["1332", "3283"], "IMDRF_E": ["1706", "2388"], "EVIDENCE": [ {"code": "1332", "quote": "fluoroscopy confirmed displacement of the ventricular lead"}, {"code": "1706", "quote": "patient experienced temporary syncope"} ], "MFR_INVESTIGATION": "Device returned and evaluated; confirmed insulation intact, mechanical dislodgement identified", "DEVICE_EXAMINED": true, "ADDRESSES_HARM": true, "FLAG": "none", "event_type": "Injury", "harm_bucket": "harm"}text
RAW FDA MAUDE MEDICAL DEVICE ADVERSE EVENT REPORT │ ▼┌─────────────────────────────────────────────────────────────┐│ RP-360/Qwen3.5-4B-mdr-narrative-grpo-codes-weighted ││ (Qwen3.5-4B Hybrid Attention + GRPO-Aligned Policy) │└─────────────────────────────────────────────────────────────┘ │ ▼ Emits Compact Codes-Only JSON (512 max tokens){ "FAILURE_MODE": "Communication failure", "IMDRF_A": ["1332"], <-- CODES ONLY "IMDRF_E": ["2388"], <-- CODES ONLY "EVIDENCE": [{"code": "1332", "quote": "THE IPG FAILED TO..."}]} │ ▼ Post-Hoc Deterministic Hydration (`enrich_record`)┌─────────────────────────────────────────────────────────────┐│ Authoritative IMDRF Taxonomy Dictionary Lookup ││ "1332" -> "Failure to Interrogate", "Problem associated..." ││ "2388" -> "Inadequate Pain Relief", "Continuing discomfort" │└─────────────────────────────────────────────────────────────┘ │ ▼ Final Enriched Output (99.6% Taxonomy Match){ "IMDRF_A": [ { "code": "1332", "term": "Failure to Interrogate", "definition": "Problem associated with the device failure to appropriately respond to signals..." } ], "IMDRF_E": [ { "code": "2388", "term": "Inadequate Pain Relief", "definition": "Continuing discomfort due to a lack or failure of a medical approach..." } ]}python
import torchfrom unsloth import FastLanguageModel model_id = "RP-360/Qwen3.5-4B-mdr-narrative-grpo-codes-weighted" model, tokenizer = FastLanguageModel.from_pretrained( model_name = model_id, max_seq_length = 2048, load_in_4bit = False,)FastLanguageModel.for_inference(model) SYSTEM_PROMPT = ( "Extract MDR information and output only valid JSON. " "For IMDRF fields, output only the codes.") narrative = """ON RECENT FOLLOW-UP, THE PATIENT EXPERIENCED SUDDEN LOSS OF STIMULATION. CHEST X-RAY REVEALED LEAD FRACTURE AT THE COSTOCLAVICULAR JUNCTION. THE PATIENT REQUIRED EMERGENCY SURGERY FOR LEAD EXTRACTION AND REPLACEMENT.""" messages = [ {"role": "system", "content": SYSTEM_PROMPT}, {"role": "user", "content": f"Extract MDR information from the following report:\n\n{narrative}"}] prompt_text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)inputs = tokenizer(prompt_text, return_tensors="pt").to("cuda") with torch.inference_mode(): outputs = model.generate( **inputs, max_new_tokens = 512, temperature = 0.1, # Low temperature recommended for deterministic extraction do_sample = False, ) response = tokenizer.decode(outputs[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True)print(response)python
import jsonimport torchfrom transformers import AutoModelForCausalLM, AutoTokenizer model_id = "RP-360/Qwen3.5-4B-mdr-narrative-grpo-codes-weighted" tokenizer = AutoTokenizer.from_pretrained(model_id)model = AutoModelForCausalLM.from_pretrained( model_id, torch_dtype = torch.bfloat16, device_map = "auto")model.eval() # Authoritative IMDRF taxonomy dictionary (subset)TAXONOMY = { "1069": ("Break", "Device separation or breakage into two or more parts."), "1332": ("Failure to Interrogate", "Problem associated with the device failure to respond to interrogation signals."), "1574": ("Failure to Deliver Stimulation", "Problem associated with the failure of the device to deliver electrical energy."), "3283": ("Wireless Communication Problem", "Problems with RF wireless technology characteristics and performance."), "2388": ("Inadequate Pain Relief", "Continuing discomfort due to a lack or failure of a medical approach to treat pain."), "1706": ("Cardiac Arrhythmia", "Any variation from the normal rate or rhythm of the heartbeat.")} def enrich_record(obj, taxonomy): """Deterministically hydrates IMDRF codes with official term and definition.""" if not obj: return obj enriched = dict(obj) enriched["IMDRF_A"] = [{"code": c, "term": taxonomy.get(c, (c, ""))[0], "definition": taxonomy.get(c, (c, ""))[1]} for c in obj.get("IMDRF_A", [])] enriched["IMDRF_E"] = [{"code": c, "term": taxonomy.get(c, (c, ""))[0], "definition": taxonomy.get(c, (c, ""))[1]} for c in obj.get("IMDRF_E", [])] enriched["EVIDENCE"] = [{"code": p["code"], "term": taxonomy.get(p["code"], (p["code"], ""))[0], "quote": p["quote"]} for p in obj.get("EVIDENCE", [])] return enriched[!WARNING] Intended Use for Post-Market Surveillance Decision Support Only This model is trained specifically to assist clinical engineers, regulatory affairs specialists, and health informatics teams in screening and organizing adverse event reports under FDA 21 CFR Part 803 and IMDRF guidelines. It is not a certified medical device and does not provide primary clinical diagnosis or patient treatment plans. All extracted codes and causality inferences must be reviewed by qualified post-market surveillance professionals before regulatory filing.
bibtex
@misc{rp360_qwen35_mdr_grpo_2026, author = {RP-360 Clinical Intelligence Team}, title = {Qwen3.5-4B MDR Narrative Intelligence: Group Relative Policy Optimization (GRPO) for Clinical Adverse Event Extraction}, year = {2026}, publisher = {Hugging Face}, howpublished = {\url{https://huggingface.co/RP-360/Qwen3.5-4B-mdr-narrative-grpo-codes-weighted}}}┌────────────────────────────────────────────────────────────────────────┐│ Composite Reward Function │├────────────────────────────────────────────────────────────────────────┤│ R = clip( ∑ (w_k · C_k) - Total_Penalties, min = -1.0, max = +1.0 )│└────────────────────────────────────────────────────────────────────────┘┌────────────────────────────────────────────────────────────────────────┐│ 7-Dimensional Clinical Reward Suite │├────────────────────────────────────────────────────────────────────────┤│ 1. IMDRF Annex A F1 Score (Device Problems) [22%] ││ 2. IMDRF Annex E F1 Score (Health Effects / Patient Harm) [22%] ││ 3. Annex Harmonic Balance (min(A, E)) [10%] ││ 4. Grounded Citation Verbatim Match Precision [12%] ││ 5. Citation Length & Semantic Quality [06%] ││ 6. Categorical Consistency (Harm Bucket, Event Type, Exam) [20%] ││ 7. Causal Chain Reasoning (Failure -> Mechanism -> Outcome) [08%] │└────────────────────────────────────────────────────────────────────────┘{ "think": "Patient presented with lead migration resulting in loss of pacing capture...", "FAILURE_MODE": "Lead displacement", "MECHANISM": "Electrode dislodgement -> loss of capture", "PATIENT_HARM": "Syncope and recurrent arrhythmia", "INTERVENTION": "Surgical revision and lead replacement", "OUTCOME": "Recovered post-reoperation", "DEVICE_ROLE": "Directly responsible", "IMDRF_A": ["1332", "3283"], "IMDRF_E": ["1706", "2388"], "EVIDENCE": [ {"code": "1332", "quote": "fluoroscopy confirmed displacement of the ventricular lead"}, {"code": "1706", "quote": "patient experienced temporary syncope"} ], "MFR_INVESTIGATION": "Device returned and evaluated; confirmed insulation intact, mechanical dislodgement identified", "DEVICE_EXAMINED": true, "ADDRESSES_HARM": true, "FLAG": "none", "event_type": "Injury", "harm_bucket": "harm"}RAW FDA MAUDE MEDICAL DEVICE ADVERSE EVENT REPORT │ ▼┌─────────────────────────────────────────────────────────────┐│ RP-360/Qwen3.5-4B-mdr-narrative-grpo-codes-weighted ││ (Qwen3.5-4B Hybrid Attention + GRPO-Aligned Policy) │└─────────────────────────────────────────────────────────────┘ │ ▼ Emits Compact Codes-Only JSON (512 max tokens){ "FAILURE_MODE": "Communication failure", "IMDRF_A": ["1332"], <-- CODES ONLY "IMDRF_E": ["2388"], <-- CODES ONLY "EVIDENCE": [{"code": "1332", "quote": "THE IPG FAILED TO..."}]} │ ▼ Post-Hoc Deterministic Hydration (`enrich_record`)┌─────────────────────────────────────────────────────────────┐│ Authoritative IMDRF Taxonomy Dictionary Lookup ││ "1332" -> "Failure to Interrogate", "Problem associated..." ││ "2388" -> "Inadequate Pain Relief", "Continuing discomfort" │└─────────────────────────────────────────────────────────────┘ │ ▼ Final Enriched Output (99.6% Taxonomy Match){ "IMDRF_A": [ { "code": "1332", "term": "Failure to Interrogate", "definition": "Problem associated with the device failure to appropriately respond to signals..." } ], "IMDRF_E": [ { "code": "2388", "term": "Inadequate Pain Relief", "definition": "Continuing discomfort due to a lack or failure of a medical approach..." } ]}import torchfrom unsloth import FastLanguageModel model_id = "RP-360/Qwen3.5-4B-mdr-narrative-grpo-codes-weighted" model, tokenizer = FastLanguageModel.from_pretrained( model_name = model_id, max_seq_length = 2048, load_in_4bit = False,)FastLanguageModel.for_inference(model) SYSTEM_PROMPT = ( "Extract MDR information and output only valid JSON. " "For IMDRF fields, output only the codes.") narrative = """ON RECENT FOLLOW-UP, THE PATIENT EXPERIENCED SUDDEN LOSS OF STIMULATION. CHEST X-RAY REVEALED LEAD FRACTURE AT THE COSTOCLAVICULAR JUNCTION. THE PATIENT REQUIRED EMERGENCY SURGERY FOR LEAD EXTRACTION AND REPLACEMENT.""" messages = [ {"role": "system", "content": SYSTEM_PROMPT}, {"role": "user", "content": f"Extract MDR information from the following report:\n\n{narrative}"}] prompt_text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)inputs = tokenizer(prompt_text, return_tensors="pt").to("cuda") with torch.inference_mode(): outputs = model.generate( **inputs, max_new_tokens = 512, temperature = 0.1, # Low temperature recommended for deterministic extraction do_sample = False, ) response = tokenizer.decode(outputs[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True)print(response)import jsonimport torchfrom transformers import AutoModelForCausalLM, AutoTokenizer model_id = "RP-360/Qwen3.5-4B-mdr-narrative-grpo-codes-weighted" tokenizer = AutoTokenizer.from_pretrained(model_id)model = AutoModelForCausalLM.from_pretrained( model_id, torch_dtype = torch.bfloat16, device_map = "auto")model.eval() # Authoritative IMDRF taxonomy dictionary (subset)TAXONOMY = { "1069": ("Break", "Device separation or breakage into two or more parts."), "1332": ("Failure to Interrogate", "Problem associated with the device failure to respond to interrogation signals."), "1574": ("Failure to Deliver Stimulation", "Problem associated with the failure of the device to deliver electrical energy."), "3283": ("Wireless Communication Problem", "Problems with RF wireless technology characteristics and performance."), "2388": ("Inadequate Pain Relief", "Continuing discomfort due to a lack or failure of a medical approach to treat pain."), "1706": ("Cardiac Arrhythmia", "Any variation from the normal rate or rhythm of the heartbeat.")} def enrich_record(obj, taxonomy): """Deterministically hydrates IMDRF codes with official term and definition.""" if not obj: return obj enriched = dict(obj) enriched["IMDRF_A"] = [{"code": c, "term": taxonomy.get(c, (c, ""))[0], "definition": taxonomy.get(c, (c, ""))[1]} for c in obj.get("IMDRF_A", [])] enriched["IMDRF_E"] = [{"code": c, "term": taxonomy.get(c, (c, ""))[0], "definition": taxonomy.get(c, (c, ""))[1]} for c in obj.get("IMDRF_E", [])] enriched["EVIDENCE"] = [{"code": p["code"], "term": taxonomy.get(p["code"], (p["code"], ""))[0], "quote": p["quote"]} for p in obj.get("EVIDENCE", [])] return enriched@misc{rp360_qwen35_mdr_grpo_2026, author = {RP-360 Clinical Intelligence Team}, title = {Qwen3.5-4B MDR Narrative Intelligence: Group Relative Policy Optimization (GRPO) for Clinical Adverse Event Extraction}, year = {2026}, publisher = {Hugging Face}, howpublished = {\url{https://huggingface.co/RP-360/Qwen3.5-4B-mdr-narrative-grpo-codes-weighted}}}