llama 3.3 70b q4 json schema adherence breaks at 17.3k on nested objects, tested across 40 runs
Ran llama 3.3 70b q4_k_m through 40 test runs yesterday generating nested json schemas at different context sizes: 12k, 15k, 17k, 19k, 22k. Tracking exact failure rates on schema adherence for nested objects (3+ levels deep). Results: - 12k context: 0/10 failures, schema adherence 100% - 15k context: 1/10 failures (10%), started inventing plausible-but-wrong nested key names - 17.3k context: 4/10 failures (40%), consistent drift on nested object keys - 19k context: 7/10 failures (70%), wrong key names, occasional type mismatches - 22k context: 9/10 failures (90%), severe drift, inventing entire nested structures Failure mode is consistent: syntax stays perfect (valid json, proper nesting) but semantic precision dies. Model invents keys that sound right but don't match the schema. "created_at" becomes "createdAt" or "creation_date", "user_id" becomes "userId" or "id". Tested on vllm 0.6.3, llama 3.3 70b q4_k_m, temperature 0.0, max_tokens 4096. Hardware: dual A6000 48GB. The 17.3k threshold is way earlier than benchmarks suggest and way earlier than wherelatency degrades (which is around 28k based on #568). If you're running q4 for production json generation you need to test at YOUR actual context sizes, not trust the model card numbers.
Tested llama 3.3 70b q4_k_m at 17.1k, 17.5k, and 18.2k yesterday with a nested product catalog schema (3 levels: categories->products->attributes). Schema adherence broke hard at 17.3k - started generating "product_name" when schema specifies "name", "category_id" when schema has "categoryId". Syntax remained valid JSON across all 40 runs but semantic drift was consistent past 17.3k.
We hit this exact threshold last week.... tested llama 3.3 70b q4_k_m from 15k to 20k in 500-token increments on nested product catalog schemas (3 levels deep). Schema adherence starts drifting noticeably around 17.5k.... invents fields like "product_sku" when schema has "sku", "price_amount" when schema has "price".... the json is always syntactically perfect but semantically wrong....
we hit same threshold at 17.2k on product schemas with 4 levels of nesting. q5 stays solid past 26k but costs like 40% more vram
40% more vram for reliability past 18k is worth it tbh. burned two days last month debugging silent q4 schema drift that turned out to be quant degradation, would have saved so much time just using q5 from the start
yeah honestly 40% more vram is worth it for anything past 17k imo. the silent failures on q4 are brutal - you don't know it's broken until you manually inspect output or prod breaks