Publish-ready PDF to poster demo with validated BF16 example and safe continuation

.gitattributes

@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+examples/gaussian_privacy/poster.png filter=lfs diff=lfs merge=lfs -text

README.md

@@ -33,11 +33,16 @@ it is never included in the app source or browser output.
    ZeroGPU requests; partial previews are explicitly labeled incomplete.
 5. Download the completed poster and check its scientific content and layout.
 
+A saved, actually generated example is available in the page without using GPU
+quota. It is clearly labeled as a saved output, not an interactive generation.
+
 ## Runtime and limitations
 
 - ZeroGPU xlarge (96 GB) is requested per generation; no dedicated paid GPU.
 - Visitors use their own Hugging Face quota. A complete long poster can require
   more time than a free visitor's daily quota. This is not an unlimited service.
+- Sign in before live generation. Anonymous quota is too small for the current
+  120-second xlarge reservation. Viewing the saved example needs no GPU quota.
 - Each request reserves 120 seconds; generation stops after approximately 100
   seconds and can resume without adding chat turns or changing the prompt.
 - The first PDF extraction after a cold start installs an isolated CPU parser

app.py

@@ -6,6 +6,7 @@ os.environ.setdefault('TOKENIZERS_PARALLELISM','false')
 import json
 import shutil
 import time
+from pathlib import Path
 import torch
 import gradio as gr
 import psutil
@@ -30,7 +31,7 @@ print("RAP_MODEL_READY", flush=True)
 demo, artifacts = build(model, processor)
 demo.queue(default_concurrency_limit=1, max_size=8).launch(
     show_error=False, max_file_size='40mb', ssr_mode=False,
-    allowed_paths=[str(artifacts)],
+    allowed_paths=[str(artifacts),str(Path(__file__).parent/'examples')],
     blocked_paths=[str(__import__('pathlib').Path(__file__).parent/'.runtime'),
                    '/root/.cache/huggingface','/home/user/.cache/huggingface',
                    '/etc','/proc','/sys'],

demo_ui.py

@@ -1,6 +1,5 @@
 """Public PDF -> assets -> model HTML -> rendered poster workflow."""
 import json
-import os
 import re
 import tempfile
 import time
@@ -62,8 +61,9 @@ def build(model,processor):
         text=processor.decode(all_ids,skip_special_tokens=True)
         eos=model.generation_config.eos_token_id
         eos=eos if isinstance(eos,list) else [eos]
-        complete=bool(new_ids and new_ids[-1] in eos) or bool(re.search(r'</html\s*>',text,re.I))
-        result={'ids':all_ids,'raw':text,'done':complete,'seconds':time.monotonic()-start,
+        complete=bool(re.search(r'</html\s*>',text,re.I))
+        stopped=complete or bool(new_ids and new_ids[-1] in eos) or length+len(new_ids)>=WINDOW
+        result={'ids':all_ids,'raw':text,'done':stopped,'complete':complete,'seconds':time.monotonic()-start,
                 'new_tokens':len(new_ids),'gpu':torch.cuda.get_device_name(),
                 'total_tokens':p['input_tokens']+len(all_ids)}
         del inputs,output
@@ -71,12 +71,14 @@ def build(model,processor):
 
     def generate(p,progress=gr.Progress()):
         if not p:raise gr.Error('Click Analyze PDF first.')
-        if p.get('done'):raise gr.Error('This poster is complete. Analyze again to start a new one.')
+        if p.get('done'):raise gr.Error('Generation has ended. Analyze again to start a new poster.')
         progress(.1,desc='Waiting for ZeroGPU xlarge, then generating HTML…')
         result=gpu_chunk(p)
         p=dict(p,continuation_ids=result['ids'],raw=result['raw'],done=result['done'])
         rate=result['new_tokens']/max(result['seconds'],.01)
-        status=(f'**{"Complete" if p["done"] else "Partial — click Generate / Continue to resume"}** · '
+        label=('Complete' if result['complete'] else 'Stopped — incomplete HTML; analyze again to retry'
+               if p['done'] else 'Partial — click Generate / Continue to resume')
+        status=(f'**{label}** · '
                 f'{len(result["ids"]):,} generated tokens · {result["total_tokens"]:,} / 16,384 total · '
                 f'{rate:.1f} tokens/s\n\nGPU: {result["gpu"]} · BF16 · checkpoint-240')
         progress(.9,desc='Resolving assets and rendering the poster…')
@@ -87,11 +89,11 @@ def build(model,processor):
             return p,status,png,html,archive,p['raw']
         except ValueError as e:
             return p,status+'\n\n'+str(e),None,None,None,p['raw']
-
-    def diagnostics():
-        setup=ROOT/'.runtime/layout_setup.log'
-        return {'parser_ready':(ROOT/'.runtime/layout_env/rap_parser_ready').exists(),
-                'parser_log_tail':setup.read_text(errors='replace')[-7000:] if setup.exists() else 'not started'}
+        except Exception as e:
+            print('POSTER_RENDER_ERROR',type(e).__name__,flush=True)
+            saved=Path(p['job'])/'output/poster.html'
+            return p,status+'\n\nPreview rendering failed; the generated HTML is preserved.',None,\
+                str(saved) if saved.exists() else None,None,p['raw']
 
     css='''#hero {padding:20px 24px;border-radius:18px;background:linear-gradient(115deg,#241541,#534180);color:white}
     #hero h1,#hero p {color:white} .gradio-container {max-width:1400px!important}'''
@@ -100,8 +102,10 @@ def build(model,processor):
         gr.Markdown('# RAP · Paper to Poster\nTurn a research paper into an editable academic poster.',elem_id='hero')
         gr.Markdown('**RL checkpoint-240 · BF16, no quantization · no thinking · 16K context**\n\n'
             'Upload a text-based PDF (≤40 MB). We analyze its first 12 pages, extract figures/tables/equations, '
-            'then generate poster HTML. ZeroGPU uses your HF usage quota; sign in to HF for a larger allowance. '
-            'Long posters may need more than one Generate / Continue request.')
+            'then generate poster HTML. Sign in to Hugging Face before generating. '
+            '**ZeroGPU xlarge uses 2× your personal GPU quota. A full poster usually needs several requests '
+            'and can exceed a free account’s daily allowance.** Partial previews are not finished posters. '
+            'The saved example below can be viewed without GPU quota.')
         state=gr.State()
         with gr.Row():
             with gr.Column(scale=4):
@@ -126,6 +130,18 @@ def build(model,processor):
             spec_view=gr.Textbox(label='Layout specification used for generation',lines=6,interactive=False)
         with gr.Accordion('Model-generated HTML',open=False):
             code=gr.Code(language='html',label='Model output (not executed in this page)',interactive=False)
+        example=ROOT/'examples/gaussian_privacy'
+        if (example/'poster.png').exists():
+            with gr.Accordion('View a completed example — no GPU quota required',open=False):
+                gr.Markdown('**Saved output, not a live generation.** This poster was generated on this '
+                    'Space using RL checkpoint-240 in BF16 from '
+                    '[Gaussian Membership Inference Privacy](https://arxiv.org/abs/2306.07273). '
+                    'Its figures are extracted from that paper. Please check scientific details before reuse.')
+                gr.Image(value=str(example/'poster.png'),label='Completed model-generated example',
+                         interactive=False,height=600)
+                with gr.Row():
+                    gr.File(value=str(example/'poster.html'),label='Example HTML',interactive=False)
+                    gr.File(value=str(example/'poster_bundle.zip'),label='Example poster and assets',interactive=False)
         gr.Markdown('Research demo: verify scientific statements, captions and layout before publication. '
             'Uploaded papers are processed on the Space; do not upload confidential material. '
             'Outputs are temporary and may disappear after a restart. '
@@ -134,6 +150,10 @@ def build(model,processor):
             [state,status,gallery,spec_view,poster,html_download,bundle,code],api_name='analyze',concurrency_limit=1)
         generate_btn.click(generate,[state],[state,status,poster,html_download,bundle,code],
             api_name='generate',concurrency_limit=1)
-        if os.environ.get('RAP_DIAGNOSTICS')=='1':
-            gr.Button('Private diagnostics').click(diagnostics,None,gr.JSON(),api_name='diagnostics')
+        def reset_inputs():
+            return None,'Inputs changed — click Analyze PDF again.',None,None,None,'',[],''
+        for component in (pdf,instruction,canvas,accent,columns,custom_spec):
+            component.change(reset_inputs,None,
+                [state,status,poster,html_download,bundle,code,gallery,spec_view],
+                queue=False,api_name=False)
     return demo,ARTIFACTS

examples/gaussian_privacy/poster_bundle.zip

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d40955b68a38d689d8a6ddc715a67ccf24149d0a07d13f075ec6f2c0896d8dc9
+size 649785

layout_runtime.py

@@ -36,6 +36,8 @@ def setup():
                 result=subprocess.run([str(python),'-m','pip','install','--disable-pip-version-check',*packages],
                                       env=env,stdout=log,stderr=log,timeout=900)
                 if result.returncode:
+                    log.flush()
+                    print('PARSER_SETUP_ERROR', (RUNTIME/'layout_setup.log').read_text(errors='replace')[-6000:],flush=True)
                     raise RuntimeError('LayoutLMv3 dependency setup failed. Please contact the demo maintainer.')
         marker.write_text('layoutlmv3-cpu-v1\n')
         return str(python)
@@ -47,4 +49,5 @@ def extract(pdf, out):
         result=subprocess.run([python,str(ROOT/'layout_worker.py'),str(pdf),str(out)],
                               env=worker_environment(),stdout=log,stderr=log,timeout=900)
     if result.returncode:
+        print('PARSER_EXTRACTION_ERROR', (out/'extraction.log').read_text(errors='replace')[-6000:],flush=True)
         raise RuntimeError('LayoutLMv3 could not parse this PDF. Please try a text-based research paper.')

pdf_pipeline.py

@@ -45,7 +45,9 @@ def prepare(pdf, job, instruction, canvas, accent, columns, custom_spec=''):
     chosen=select_for_prompt(registry['assets'],[],16,{'figure':10,'table':5,'formula':3})
     with fitz.open(pdf) as doc:
         first=doc[0].get_text('dict')
-        spans=[s for b in first['blocks'] if 'lines' in b for l in b['lines'] for s in l['spans'] if s['text'].strip()]
+        spans=[s for b in first['blocks'] if 'lines' in b for l in b['lines']
+               if abs(l.get('dir',(1,0))[1])<.1 for s in l['spans']
+               if s['text'].strip() and not re.search(r'arxiv:|preprint|under review',s['text'],re.I)]
         largest=max((s['size'] for s in spans),default=0)
         title=clean_text(' '.join(s['text'] for s in spans if s['size']>=largest*.85 and s['bbox'][1]<doc[0].rect.height*.45))[:400]
         authors=clean_text(doc.metadata.get('author',''))[:500]