Conversation
…anity check - Add MODEL_FAMILIES config table with per-model API params and server flags - Add getModelApiParams() helper to inject reasoning_effort:none for Mistral - Add delta.thinking fallback in streaming loop to capture thinking tokens - Add streaming sanity check before benchmark run (detects empty-token loops) - Add test-model-config.cjs with 17 unit tests for model detection logic
Feature/mistral thinking fix
Add Mistral Small 4 (119B, IQ1_M + Q2_K_XL), NVIDIA Nemotron-3-Nano (4B + 30B), Liquid LFM2 (1.2B + 24B), Qwen3.5-9B BF16, and Qwen3.5-27B Q8_K_XL to the benchmark paper. Key updates: - Abstract: 7→16 models, 5 families, best local now 95.8% - Models Under Test table: grouped by family, 16 rows - Overall Scorecard: full 16-model ranking - Key Finding 3: quantization precision > parameter count - Conclusion: Qwen3.5-27B Q8 at 95.8%, Mistral-119B at 89.6%
Add Nemotron and LFM2 model families to MODEL_FAMILIES with minTemperature: 1.0 — these models reject temperature < 1.0 with HTTP 400. The benchmark now clamps temperature to the family minimum before sending the request. - Refactor getModelApiParams → getModelFamily (returns full config) - Add resolveTemperature logic in llmCall params builder - Update test-model-config.cjs: 27 tests including temperature clamp - Fix Mistral serverFlags to match current llm-server-manager.cjs
codeCraft-Ritik
left a comment
codeCraft-Ritik
left a comment
There was a problem hiding this comment.
Great work! The implementation is clean and easy to understand
llama-server with Qwen3.5/Claude-distilled models outputs thinking as 'Let me analyze...' plain text in delta.content (no <think> tags, no separate reasoning field). The JSON-expect abort was firing at 50 chars, killing the request after 8-10 tokens before the model could output actual JSON. Changes: - Raised JSON content check threshold from 50 to 200 chars - Strip common plain-text reasoning prefixes before checking - Only abort if 200+ chars of non-JSON, non-reasoning content
For JSON-expected tests, delta.content text arriving before any
JSON start ({/[) is now routed to reasoningContent, not content.
This handles llama-server with Qwen3.5-Claude models where thinking
appears as plain text in delta.content without <think> tags.
The model's thinking is logged/shown but NOT evaluated as output.
Only the actual JSON content (after the first {/[) is treated as
the model's response for test evaluation.
- Docker-only deployment for all platforms (Linux, macOS, Windows) - Docker Desktop 4.35+ USB/IP for macOS/Windows USB passthrough - YOLO26n Edge TPU model (INT8, 320x320, ~4ms inference) - pycoral-based inference with CPU fallback - JSONL stdin/stdout protocol (same as yolo-detection-2026) - deploy.sh/deploy.bat for autonomous Docker image build - Colab/Kaggle compilation script for Edge TPU model - TPU device selector and clock speed config
- Docker deployment using official openvino/ubuntu22_runtime image - Supports Intel NCS2 (MYRIAD), Intel GPU (iGPU/Arc), and CPU - AUTO device selector lets OpenVINO pick best available - FP16/INT8/FP32 precision options - YOLO26n with Ultralytics OpenVINO backend - JSONL stdin/stdout protocol (same as yolo-detection-2026) - Colab script for model export (runs on any platform)
… update OpenVINO detect.py: - Add file_read timing metric (matches Coral TPU) - Add frame-not-found guard in main loop (empty detections response) - Add invalid JSON log message instead of silent continue OpenVINO SKILL.md: - Add description fields to all parameters - Add Platform Setup (Linux/macOS/Windows) section - Add Model section with compile instructions - Add Bounding Box Format section OpenVINO deploy.sh: - Add find_docker() function pattern - Add exit code 2 for partial success (CPU-only) - Add architecture to platform progress event - Add accelerator_found field in complete event OpenVINO deploy.bat: - Add Docker version reporting - Add device probe result checking New: scripts/compile_model.py - Local model export (--model, --size, --precision, --output) - FP16/INT8/FP32 via YOLO.export(format=openvino) README.md: - Add Coral TPU and OpenVINO to Skill Catalog (🧪 Testing) - Add Detection & Segmentation Skills architecture section - Add mermaid diagram showing native vs Docker detection paths - Add LLM-Assisted Skill Installation explanation
feat: add YOLO 2026 Coral TPU detection skill (Docker-based)
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
No description provided.