Streamline model card and refresh chart

README.md

@@ -14,23 +14,19 @@ tags:
 
 # nyx
 
-nyx is a compact 4B decision model for Choice, Score, and Noul workloads. The release is one 3.29 GB GGUF that runs with llama.cpp on CPU, Apple Metal, NVIDIA CUDA, and AMD HIP.
+nyx is a compact 4B decision model for Choice, Score, and Noul workloads. The release is one 3.29 GB GGUF for llama.cpp on CPU, Apple Metal, NVIDIA CUDA, and AMD HIP.
 
-![Animated light-theme chart comparing the nyx reference and Jev across six classification benchmarks](assets/nyx-benchmarks.svg)
+![Animated green-and-white chart comparing nyx and Jev across six classification benchmarks](assets/nyx-benchmarks.svg)
 
 ## Model
 
 | Property | Value |
 |---|---|
-| Base | [`Qwen/Qwen3.5-4B`](https://huggingface.co/Qwen/Qwen3.5-4B) |
-| Base revision | `851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a` |
+| Base | [Qwen/Qwen3.5-4B](https://huggingface.co/Qwen/Qwen3.5-4B) |
 | File | `nyx.gguf` |
-| Size | 3,288,582,784 bytes (3.29 GB / 3.06 GiB) |
-| Quantization | Importance-matrix IQ4_XS with selected Q5_K/Q6_K tensors |
+| Size | 3.29 GB / 3.06 GiB |
+| Format | GGUF, importance-matrix IQ4_XS with selected Q5_K/Q6_K tensors |
 | SHA-256 | `39d64fb19863cfc067e2411fcb2a0444113dda13798b15de5617736d91914359` |
-| Runtime | Pinned llama.cpp build in [`devanshbatham/nyx`](https://github.com/devanshbatham/nyx) |
-
-The architecture, tokenizer, and base weights originate from Qwen. nyx adds decision-task adaptation, task-vector composition, quantization, and a production API integration.
 
 ## Run
 
@@ -38,7 +34,6 @@ The architecture, tokenizer, and base weights originate from Qwen. nyx adds deci
 git clone https://github.com/devanshbatham/nyx.git /opt/nyx
 hf download devanshbatham/nyx --local-dir /opt/nyx-model
 
-# auto-detects CPU, Metal, CUDA, or HIP
 /opt/nyx/scripts/build-llama-cpp.sh /opt/llama.cpp
 
 export LLAMA_SERVER_BIN=/opt/llama.cpp/build/bin/llama-server
@@ -46,13 +41,13 @@ export NYX_GGUF_PATH=/opt/nyx-model/nyx.gguf
 /opt/nyx/scripts/serve-model.sh
 ```
 
-Budget at least 8 GB of free RAM, VRAM, or unified memory for one 4,096-token slot and 6 GB of free disk. More context and concurrent slots require more memory. The model backend binds to unauthenticated loopback port `30000`; do not expose it publicly.
+Allow 8 GB of free RAM, VRAM, or unified memory and 6 GB of disk for one 4,096-token slot. Larger contexts and concurrent slots need more memory.
 
-For the authenticated `POST /v1/systemone` API and typed Python/TypeScript clients, use [`github.com/devanshbatham/nyx`](https://github.com/devanshbatham/nyx). The gateway implements the TypeSafe System One wire contract for Choice, Score, Noul, model discovery, bearer authentication, and request IDs.
+The production API and typed Python/TypeScript clients are in [devanshbatham/nyx](https://github.com/devanshbatham/nyx).
 
-## Evaluation
+## Benchmarks
 
-The public comparison below is for the pre-quantization nyx reference. The pretrained Qwen base and nyx reference used the same local runtime; all three systems received the same 2,277 frozen requests.
+Accuracy on 2,277 frozen classification requests:
 
 | Benchmark | Qwen3.5-4B base | nyx reference | Jev 1.13 |
 |---|---:|---:|---:|
@@ -67,29 +62,8 @@ The public comparison below is for the pre-quantization nyx reference. The pretr
 | Round trip, median / p95 | 54.85 / 67.96 ms | 55.56 / 72.77 ms | 182.33 / 245.12 ms |
 | Throughput | 17.44 req/s | 17.15 req/s | 42.43 req/s |
 
-Local timings used one AMD Instinct MI325X, one execution slot, Q8 KV cache, Flash Attention, and a 4,096-token context. Jev used remote HTTPS at concurrency 8, so its timing is descriptive rather than a hardware-controlled speed comparison.
-
-### Compact-release fidelity
-
-The released GGUF was compared with the 8.42 GB reference on a separate frozen 1,024-case set.
-
-| Metric | Reference | Released GGUF |
-|---|---:|---:|
-| Correct | 873/1,024 | 874/1,024 |
-| Accuracy | 85.25% | 85.35% |
-| Decision agreement | — | 98.73% (1,011/1,024) |
-| Decision disagreements | — | 13 |
-
-The compact model gained seven decisions and lost six relative to the reference. It did not meet the experiment's original 99.5% agreement target; publication followed an explicit decision to accept that tradeoff. See the [evaluation record](benchmarks/random-six-report.md).
-
-## Limitations
-
-- Optimized for structured classification and decision workloads; general assistant quality was not established.
-- The released GGUF has not been rerun on the full public 2,277-case suite.
-- NVIDIA and Apple performance and cross-vendor decision parity have not been measured.
-- Public-benchmark pretraining overlap is unknown.
-- Outputs can be incorrect or biased. Validate consequential decisions independently.
+Local Qwen and nyx latency: MI325X, concurrency 1. Jev latency: remote HTTPS, concurrency 8. The published GGUF matched the nyx reference on 98.73% of a separate 1,024-case fidelity set. [Reproduce the nyx/Jev benchmark](https://github.com/devanshbatham/nyx/tree/main/benchmarks).
 
-## License and attribution
+## License
 
-Apache License 2.0. nyx is derived from [`Qwen/Qwen3.5-4B`](https://huggingface.co/Qwen/Qwen3.5-4B), also distributed under Apache-2.0. Qwen and Alibaba Cloud are credited as the original model authors. This project is not endorsed by or affiliated with Qwen, Alibaba Cloud, TypeSafe, or Jev.
+Apache-2.0. nyx is derived from [Qwen/Qwen3.5-4B](https://huggingface.co/Qwen/Qwen3.5-4B); Qwen and Alibaba Cloud are credited as the original model authors. This project is not affiliated with Qwen, Alibaba Cloud, TypeSafe, or Jev.

benchmarks/random-six-report.md

@@ -1,71 +0,0 @@
-# Six public classification benchmarks
-
-These results describe the 8.42 GB pre-quantization nyx reference, not the 3.29 GB release GGUF. The compact file was selected using the separate fidelity evaluation reported below.
-
-Frozen paired cohort: 2,277 cases; case SHA-256 `046a38c24747fd724003aca7e5f1d1b6e148d599092222361c0717ecff84ecad`; every failure counts as wrong.
-
-| Dataset | Cases | Qwen3.5-4B base | nyx | Jev 1.13 |
-|---|---:|---:|---:|---:|
-| AG News | 400 | 87.75% (351/400) | 85.75% (343/400) | 86.50% (346/400) |
-| TREC coarse | 500 | 86.60% (433/500) | 84.20% (421/500) | 93.00% (465/500) |
-| CoLA | 400 | 77.75% (311/400) | 81.50% (326/400) | 77.75% (311/400) |
-| RTE | 277 | 84.12% (233/277) | 88.09% (244/277) | 91.34% (253/277) |
-| Emotion | 300 | 47.67% (143/300) | 46.33% (139/300) | 47.00% (141/300) |
-| IMDb | 400 | 95.75% (383/400) | 95.75% (383/400) | 97.00% (388/400) |
-| Dataset-macro accuracy | 2,277 | 79.94% | 80.27% | 82.10% |
-| Dataset-macro F1 | 2,277 | 79.83% | 80.48% | 81.85% |
-| Pooled micro accuracy | 2,277 | 81.42% (1,854/2,277) | 81.51% (1,856/2,277) | 83.62% (1,904/2,277) |
-
-Jev-only versus nyx-only correct cases were 121 to 73; the exact two-sided McNemar p-value is `0.0006995`. nyx-only versus Qwen-only correct cases were 112 to 110 (`p=0.9465`).
-
-## Runtime
-
-The release profile uses one llama.cpp execution slot because it was the fastest measured configuration for this latency-oriented workload.
-
-| Measurement | Result |
-|---|---:|
-| Reference model file | 8,424,393,056 bytes |
-| Measured VRAM | 9,927,208,960 bytes |
-| Round trip, median | 55.56 ms |
-| Round trip, p95 | 72.77 ms |
-| Round trip, mean | 57.92 ms |
-| Full-run wall time | 132,780.73 ms |
-| Sustained throughput | 17.15 requests/s |
-
-The nyx timing is direct loopback HTTP to llama.cpp on one AMD Instinct MI325X, with full GPU offload, Flash Attention, Q8 KV cache, 4,096 tokens per slot, and client/server concurrency 1. It excludes the authenticated gateway and external network. The Qwen base used the identical local profile and measured 54.85 ms median, 67.96 ms p95, and 17.44 requests/s.
-
-The previously observed Jev remote-HTTPS run used client concurrency 8 and reported 182.33 ms median, 245.12 ms p95, and 42.43 requests/s. Different hardware, network topology, and concurrency make that timing descriptive rather than a controlled speed comparison.
-
-## Batching check
-
-The first 120 frozen cases were repeated at three concurrency levels. The server had the same number of 4,096-token slots as the client concurrency.
-
-| Concurrency | Correct | Median / p95 | Throughput |
-|---:|---:|---:|---:|
-| 1 | 94/120 | 57.20 / 89.01 ms | 15.79 requests/s |
-| 2 | 94/120 | 162.80 / 219.25 ms | 11.77 requests/s |
-| 8 | 94/120 | 606.22 / 887.68 ms | 11.61 requests/s |
-
-All 120 class decisions were identical across the three runs. Batching was rejected because it reduced throughput and increased latency.
-
-## Protocol and limits
-
-- Dataset selection, sampled examples, request order, and labels were frozen before provider scoring.
-- AG News, CoLA, Emotion, and IMDb are fixed class-stratified samples; TREC test and RTE validation are complete.
-- Each request contained one Choice question. Both systems completed 2,277/2,277 cases with zero errors.
-- Public-benchmark pretraining overlap is unknown. This finite suite does not establish universal classification quality.
-- The NVIDIA RTX 4000 Ada target still requires an on-device latency and class-decision parity run before production rollout.
-
-## Compact release fidelity
-
-The released importance-matrix IQ4_XS candidate was evaluated against the reference on a frozen 1,024-case selection spanning Banking77, BoolQ, CLINC150, code-defect annotation, commit type, DBpedia14, email spam, SNLI, structured rules, and tweet sentiment.
-
-| Metric | Reference | Released GGUF |
-|---|---:|---:|
-| File size | 8,424,393,056 bytes | 3,288,582,784 bytes |
-| Correct | 873/1,024 | 874/1,024 |
-| Accuracy | 85.25% | 85.35% |
-| Decision agreement | — | 98.73% (1,011/1,024) |
-| Decision disagreements | — | 13 |
-
-The released candidate gained seven decisions and lost six relative to the reference. This is fidelity evidence, not a fresh public-benchmark run. The candidate did not meet the experiment's original 99.5% agreement target and was released only after that threshold was explicitly waived.

release-manifest.json

@@ -14,20 +14,16 @@
       "sha256": "a7c85805e2a02824e60d752e1d543649918c081b45b150615a926ea2f79f2e91"
     },
     "README.md": {
-      "bytes": 4518,
-      "sha256": "dcb056c79af04ccbc3662af1bee3a480885cac0451d3322f2399bfca6cdead35"
+      "bytes": 2568,
+      "sha256": "8a64b70816f33dd4997b1b815ab267ad5e2192f3401425d1cadc9dc69e6fca5a"
     },
     "assets/IOSEVKA-OFL.md": {
       "bytes": 4674,
       "sha256": "4fbfa00e9fb9e9a774ab4c099bc6e8b2ac1a46f6492a2c260f958b004d172da6"
     },
     "assets/nyx-benchmarks.svg": {
-      "bytes": 121931,
-      "sha256": "d0096f7d804d89f1cebec9e1e966cb05c855382b9842f5741cb162e3f7b884d5"
-    },
-    "benchmarks/random-six-report.md": {
-      "bytes": 4372,
-      "sha256": "68f53d0f47f25a3cf92fe73be76537164270b18b10d374e2247456f7daf3036f"
+      "bytes": 121677,
+      "sha256": "3d950fd5ce930e3309c444513c45e2748c7716029fca4bda4b999a8b23fce1fa"
     },
     "nyx.gguf": {
       "bytes": 3288582784,
@@ -40,5 +36,5 @@
   },
   "format_version": 1,
   "repository": "devanshbatham/nyx",
-  "total_bytes": 3288731957
+  "total_bytes": 3288725381
 }