--- license: apache-2.0 language: - en library_name: pytorch pipeline_tag: text-generation tags: - code - recurrent - peer - mixture-of-experts - evalplus --- # DNA-DiskChat-2B-PEER-v25-Code Recurrent PEER code model. This card is generated from measured JSON. Missing values remain **not measured**. ## Model Details | Field | Value | |---|---:| | Logical parameters | 1999254018 | | Active parameters/token | 40189952 | | Checkpoint bytes | 195637409 | | Effective bits/parameter | 0.783 | | Compression validation CE delta | 4.952 | | Compression quality gate (`<= +0.01`) | False | | License | apache-2.0 | ## Code Benchmarks EvalPlus results use greedy pass@1 generation. Base and Plus tests are reported separately. | Benchmark | Base pass@1 | Plus pass@1 | Tasks | |---|---:|---:|---:| | HumanEval | 0.00% | 0.00% | 164 | | MBPP | 0.00% | 0.00% | 378 | ## Training Throughput | Metric | Value | |---|---:| | Mean tokens/s | 55,503.942 | | Median tokens/s | 55,655.991 | | Measured windows | 24611 | ![Training throughput](charts/training_tok_s.png) ![EvalPlus code benchmarks](charts/evalplus.png) ## NVMe `O_DIRECT` Benchmark This transport benchmark measures random compressed-page lookups, not model tokens. `O_DIRECT` bypasses Linux page cache but not NVMe controller caches. | Metric | Value | |---|---:| | Page lookups/s | 9,276.357 | | p50 latency | 58.041 us | | p95 latency | 206.685 us | | Requested bytes/lookup | 4096 | | Useful bytes/lookup | 3,000.000 | | Direct NVMe qualified | False | | Filesystem | overlayfs | ## Intended Use Research on compact recurrent code generation and sparse expert routing. Generated code is untrusted; review it and execute only in a sandbox. ## Limitations - Benchmark scores do not establish security, correctness, or production readiness. - Missing metrics were not measured and are not estimated from plans or earlier model versions. - EvalPlus execution runs untrusted model-generated Python and requires isolation. - Page-lookup/s is not inference tok/s. Token speed requires integrated router, decoder, and expert compute. ## Reproducibility Canonical report schema version: `1`. Generated at: `2026-08-02T00:54:22.028242+00:00`.