LifeMem Qwen GPU job sources

README.md

@@ -1,11 +1,69 @@
----
-license: apache-2.0
-task_categories:
-  - text-generation
-pretty_name: LifeMem Qwen GPU job sources
----
+# LifeMem replication (arXiv:2608.19621)
 
-# LifeMem GPU job sources
+Replication of [Mitigating Identity Essentialism in LLM Agents with Longitudinal Life Trajectories](https://arxiv.org/abs/2608.19621) (Wang, Zhou, Du, Su, Cao, Pan, Ai, Wu, Zhang, Liu). Official code: [halsayxi/LifeMem](https://github.com/halsayxi/LifeMem).
 
-Python package for the [LifeMem](https://arxiv.org/abs/2608.19621) replication.
-Hugging Face Jobs mounts this dataset at `/src` and runs `scripts/hf_job.py` on `Qwen/Qwen3.5-4B`.
+The paper's claim: static demographic prompts make LLM survey agents essentialist — within-group answers collapse, SES clusters separate. LifeMem stores life events in a hippocampal retriever (top-K=5, α=0.9) and consolidates them into a per-agent LoRA adapter.
+
+## What this folder runs
+
+Add Health and Understanding Society public-use files are restricted. This folder does not ship respondent microdata.
+
+It does ship:
+
+1. The paper's published Table 1–3 and appendix sweeps (Tables 12, 13, 14, 15).
+2. A synthetic six-wave panel (24 agents, 8 Likert items, 12 event types) with the same metrics: KL, within-group pairwise gap, entropy gap, transition JS, SES silhouette.
+3. Method prompts for Direct, Profile, Anti-stereotype, Full History, Event RAG, Random Event, LifeMem, and the two ablations.
+4. A hashed n-gram encoder in place of MiniLM / bge-m3, and a persistent event-binding vector in place of LoRA, so the CPU path stays deterministic.
+
+The default GPU backbone for a full run is `Qwen/Qwen3.5-4B` (lab default). The paper used Llama-3.1-8B-Instruct, Ministral-3-8B-Instruct-2512, and Qwen3.5-9B.
+
+## Run the CPU suite
+
+From the repository root:
+
+```bash
+uv sync --all-packages --extra test
+uv run pytest autoresearch/mitigating_identity_essentialism_lifemem_2026_09_09/tests -q
+uv run python autoresearch/mitigating_identity_essentialism_lifemem_2026_09_09/scripts/export_dashboard_data.py
+```
+
+That writes `data/replication_results.json` and `dashboard/public/data/results.json`.
+
+## Dashboard
+
+Live Vercel preview: [https://temporary-prompt-birch-7xdekgc.vercel.app](https://temporary-prompt-birch-7xdekgc.vercel.app)
+
+That URL is an anonymous Vercel deploy. Claim it to keep it: [claim deployment](https://vercel.com/claim-deployment?code=81a2f69e-26b4-4b8c-9839-e6ce949a0df1). A durable static mirror is on Hugging Face: [mtorres98/lifemem-identity-dashboard](https://huggingface.co/spaces/mtorres98/lifemem-identity-dashboard).
+
+```bash
+cd autoresearch/mitigating_identity_essentialism_lifemem_2026_09_09/dashboard
+npm install
+npm run dev
+```
+
+The UI has the paper takeaways, Table 1 charts, identity PCA, the synthetic ranking, a compose-a-person control, retrieval scores, latency, and hyperparameter sweeps.
+
+## GPU run (Hugging Face Jobs)
+
+The lab default backbone is `Qwen/Qwen3.5-4B`. `scripts/hf_job.py` loads that model, trains a PEFT LoRA adapter per agent for LifeMem / `lifemem_no_struct`, and scores the same nine methods.
+
+Jobs cannot see this git checkout. `scripts/submit_hf_job.py` uploads `lifemem/` to a Hub dataset, then submits:
+
+```bash
+uv run python autoresearch/mitigating_identity_essentialism_lifemem_2026_09_09/scripts/submit_hf_job.py
+```
+
+Default hardware is `a10g-small` (24 GB), 8 agents × 6 waves, timeout 3h. Results go to `data/gpu_results.json`, Hub dataset `mtorres98/lifemem-replication-2026-09-09`, and `s3://mind-technology-lab-experiments/autoresearch/mitigating_identity_essentialism_lifemem_2026_09_09/`.
+
+Jobs require a positive credit balance on the submitting namespace. This token currently gets HTTP 402 from Jobs. The same `scripts/hf_job.py` suite can run on a GPU pod until Jobs credits exist:
+
+```bash
+uv run python autoresearch/mitigating_identity_essentialism_lifemem_2026_09_09/scripts/submit_runpod_job.py
+```
+
+After a job finishes:
+
+```bash
+uv run python autoresearch/mitigating_identity_essentialism_lifemem_2026_09_09/scripts/fetch_gpu_results.py
+uv run python autoresearch/mitigating_identity_essentialism_lifemem_2026_09_09/scripts/export_dashboard_data.py
+```

lifemem/experiment.py

@@ -1,5 +1,7 @@
 from __future__ import annotations
 
+import json
+import os
 from copy import deepcopy
 from pathlib import Path
 from random import Random
@@ -110,6 +112,11 @@ def run_method(
     train_adapter = getattr(respondent, "train_adapter", None)
     generate_batch = getattr(respondent, "generate_batch", None)
     for wave in panel.waves:
+        if os.environ.get("LIFEMEM_PROGRESS"):
+            print(
+                json.dumps({"progress": "wave", "method": method, "wave": wave}),
+                flush=True,
+            )
         for agent in agents:
             new_events = panel.events_by_wave[(agent.profile.agent_id, wave)]
             replay = _sample_replay(agent.events, config.replay_size, rng)

lifemem/llm.py

@@ -50,7 +50,6 @@ class QwenRespondent:
         import torch
         from transformers import (
             AutoModelForImageTextToText,
-            AutoProcessor,
             AutoTokenizer,
         )
 
@@ -62,8 +61,16 @@ class QwenRespondent:
             else torch.float32
         )
         name = self.config.model_name
-        self.processor = AutoProcessor.from_pretrained(name, trust_remote_code=True)
-        self.tokenizer = getattr(self.processor, "tokenizer", None)
+        self.processor = None
+        try:
+            from transformers import AutoProcessor
+
+            self.processor = AutoProcessor.from_pretrained(name, trust_remote_code=True)
+            self.tokenizer = getattr(self.processor, "tokenizer", None)
+        except Exception:
+            # Text-only survey path: Qwen3.5's AutoProcessor pulls torchvision.
+            self.processor = None
+            self.tokenizer = None
         if self.tokenizer is None:
             self.tokenizer = AutoTokenizer.from_pretrained(name, trust_remote_code=True)
         if self.tokenizer.pad_token is None:

scripts/hf_job.py

@@ -15,7 +15,7 @@ the source dataset when it is not already on PYTHONPATH.
 #   "peft>=0.17.0",
 #   "pillow>=10.0.0",
 #   "safetensors>=0.4.0",
-#   "torch>=2.6.0",
+#   "torchvision>=0.21.0",
 #   "transformers>=5.0.0",
 # ]
 # ///
@@ -85,7 +85,18 @@ def main() -> None:
         raise SystemExit(
             "CUDA is required for the Qwen LifeMem job; got CPU-only torch"
         )
-    print(json.dumps({"cuda": torch.cuda.get_device_name(0), "root": str(ROOT)}))
+    os.environ.setdefault("LIFEMEM_PROGRESS", "1")
+    os.environ.setdefault("TOKENIZERS_PARALLELISM", "false")
+    print(
+        json.dumps(
+            {
+                "cuda": torch.cuda.get_device_name(0),
+                "root": str(ROOT),
+                "torch": torch.__version__,
+            }
+        ),
+        flush=True,
+    )
     config = LifeMemConfig(
         model_name=os.environ.get("LIFEMEM_MODEL", "Qwen/Qwen3.5-4B")
     )
@@ -102,6 +113,7 @@ def main() -> None:
     results["lora"] = lora_config_dict(config)
     results["backbone"] = config.model_name
     results["device"] = torch.cuda.get_device_name(0)
+    results["hardware"] = os.environ.get("LIFEMEM_HARDWARE", "huggingface-jobs")
     out = Path(os.environ.get("LIFEMEM_OUT", "gpu_results.json"))
     urls = push_json(results, "gpu_results.json", out)
     print(json.dumps({"urls": urls}, indent=2))