What is this defeatist thinkng from the model???
It is ridiculous. No matter how much context I allocate, or how much output allowance, the model complains about the amount of work. It is full of doubt that it can do the job. I have tried to adjust the system instructions to make it more direct, avoid nuance and it continues complaining about the amount of work.
Don't you advertise that the context size is one(1) million? I am using half -ctx 500000 and -n 400000.
<think>
We need answer only XML document. Need process huge input. Need produce comprehensive ontology. We need decide scope. We have many documents about poorly differentiated thyroid carcinoma (PDTC), studies, case reports, etc. Need represent all substantive info. This is enormous. We need create ontology with classes, individuals, properties, assertions. We must not include packaging IDs. Need preserve source URLs, pages, dates, etc. Need cover all supplied source content. We have token budget? We need output likely large. We need be systematic but concise. We can create ontology with classes for Document, Study, Patient, Disease, Treatment, Outcome, Measurement, etc. But must cover all facts. Given huge input, impossible to represent every sentence in full detail within output? We need do best. We can create a structured ontology with key entities and facts. The instructionsays comprehensive source coverage. We need include all documents, their metadata, authors, affiliations, abstracts, results, tables, etc. That's massive. But we can represent at a reasonable level: each document as SourceDocument with URL, PMID, DOI, journal, volume, issue, pages, publication dates, authors, affiliations. Then key findings as claims/measurements. We can create individuals for studies, patients, treatments, outcomes. But there are many numbers. We need include all numeric values? The instruction says preserve every value's subject, meaning, units, denominator, baseline, precision, qualification. That's huge. We may need to include many measurements. But output length limited. We need prioritize. The evaluation likely checks for coverage of key facts, not every number? But prompt says exhaustive. We need produce as much as possible.
<think>
We need answer only XML. Need process all supplied source content into one complete ontology. This is huge. Need produce comprehensive but manageable. We need extract every supplied item? Impossible fully in token budget, but need best effort. We need output one XML document. We must include all documents, facts, entities, events, times, measurements, relationships, sources. Need not include every reference? "ALL supplied source content" includes many references. We need represent all info. But token budget ~? We have 768 juice, final can be long. Need craft ontology. We can create classes/properties and individuals for documents, passages, studies, patients, treatments, measurements, times, links. Need preserve source URLs, PMIDs, DOIs, dates, stats. We can encode many facts as individuals with labels and data properties. Need not be perfect but comprehensive.
Non-stop worrying about token budgets, saying "This is huge," "best efforts," etc...
It is basically getting ready to do a bad, incomplete job but is justifying it by claiming it is too much. I have never seen another model complain so much about the size of the work.
And the "huge" amount of work it complains about is less than 75000 tokens, including instructions and 50000 data. Why is this happening with 1 million token context model? It appears to be pre-trained to do much less than the best work.
The "many documents" it claims, are only 7 documents; 4 pages + 9 pages + 1 + 1 + 1 + 1 + 1 web pages (abstracts)
"Need not be perfect?" WTH?
The only issue is that it thinks so much - to make this 3d marble maze in html+javascript it took 65,000 tokens and its not perfect. The maze doesn't allow the ball to roll and move through the maze.
At 14 t/s it takes hours to complete the task. And this is using the Q3 model. The Q6 was 4 t/s.
So, I will stick with Qwen 3.8 27B and the other Qwen models.