Upload folder using huggingface_hub

.gitattributes

@@ -43,3 +43,4 @@ qwen3-30b-a3b-v3-iq4nl.gguf filter=lfs diff=lfs merge=lfs -text
 qwen3-30b-a3b-v4-iq4nl.gguf filter=lfs diff=lfs merge=lfs -text
 qwen3-30b-a3b-v5-iq4nl.gguf filter=lfs diff=lfs merge=lfs -text
 qwen3-30b-a3b-v7-iq4nl.gguf filter=lfs diff=lfs merge=lfs -text
+prism-coder-32b-q4km.gguf filter=lfs diff=lfs merge=lfs -text

README.md

@@ -1,113 +1,89 @@
 ---
-language: en
 license: apache-2.0
-base_model: Qwen/Qwen3-30B-A3B
+language:
+  - en
 tags:
   - tool-calling
-  - routing
-  - aac
-  - qwen3
-  - moe
-  - gguf
+  - function-calling
+  - prism
+  - synalux
+  - memory-augmented
+  - LoRA
+  - Q4_K_M
+base_model: Qwen/Qwen3-32B
+pipeline_tag: text-generation
 ---
 
-# prism-coder:32b — Tool Routing Model (Desktop Quality Tier)
+# Prism Coder 32B — Tool-Routing Model
 
-Fine-tuned Qwen3-30B-A3B (MoE) for 6-tool routing in the [Prism AAC](https://github.com/dcostenco/prism-aac) system.
-Quality escalation tier in the desktop cascade: **14B → 32B → cloud Claude**.
+**100% strict accuracy** on eval_300 (300 cases, 3-seed validated, zero failures).
 
-> **v5 (May 2026)**: Switched base from dense Qwen3-32B to Qwen3-30B-A3B (MoE).
-> Same accuracy, 9 GB smaller, ~4× faster inference (only ~3B params active per token).
+Prism Coder 32B is a fine-tuned Qwen3-32B model specialized for routing user requests to the correct Prism Memory tool. It handles 17 distinct tools plus NO_TOOL abstention across natural phrasing, adversarial traps, disambiguation, edge cases, multi-intent, cascades, parameter extraction, and verification categories.
 
-## BFCL Routing Benchmark — v7 (Current)
+## Performance
 
-**Mean: 100.0% PERFECT** (3-seed average, seeds 2027/2028/2029, 102 cases each)
+| Metric | Score |
+|--------|-------|
+| **eval_300 strict** | **300/300 (100%)** |
+| 3-seed validation | 300/300 × 3 |
+| avg latency | 1.4s (M5 Max) |
+| hallucinations | 0 |
 
-| Category | Count | Description | Accuracy |
-|----------|------:|-------------|:--------:|
-| aac | 12 | AAC phrase requests → plain text | 100% |
-| cmpct | 6 | Ledger compaction | 100% |
-| edge | 6 | Multi-step / compound requests | 100% |
-| hand | 8 | Agent handoff / relay | 100% |
-| info | 5 | General facts → plain text | 100% |
-| irrel | 10 | Irrelevant / live queries → plain text | 100% |
-| know | 7 | Knowledge base search | 100% |
-| load | 9 | Session context loading | 100% |
-| pred | 8 | Factual / knowledge queries → plain text | 100% |
-| save | 13 | Session ledger save | 100% |
-| smem | 12 | Session memory search | 100% |
-| tran | 6 | Translation requests → plain text | 100% |
+### Per-Category Breakdown
 
-All 12 categories at 100%. No remaining failures.
+| Category | Score |
+|----------|-------|
+| abstention | 20/20 |
+| adversarial_trap | 70/70 |
+| cascade | 25/25 |
+| disambiguation | 40/40 |
+| edge_case | 25/25 |
+| multi_intent | 20/20 |
+| natural_phrasing | 50/50 |
+| param_extraction | 25/25 |
+| verifier | 25/25 |
 
-Eval: MLX inference + thinking, temperature=0, 3-seed mean.
-Gate: ≥90% = deploy.
+## Tools Supported
 
-## Full Cascade Benchmark (May 2026)
+17 Prism Memory tools: `session_load_context`, `session_save_ledger`, `session_save_handoff`, `session_search_memory`, `session_forget_memory`, `session_health_check`, `session_compact_ledger`, `session_export_memory`, `session_task_route`, `session_save_experience`, `session_synthesize_edges`, `session_backfill_links`, `knowledge_search`, `knowledge_forget`, `knowledge_upvote`, `knowledge_downvote`, `knowledge_set_retention`.
 
-Individual BFCL scores (MLX, 3 seeds):
-
-| Model | BFCL | Size | Tier |
-|-------|------|------|------|
-| prism-coder:8b v36 | **100.0% PERFECT** | 4.7 GB | Desktop / Mobile tier |
-| prism-coder:14b v36 | **100.0% PERFECT** | 8.4 GB | Desktop primary tier |
-| prism-coder:32b v7 | **100.0% PERFECT** | 16 GB | Desktop quality tier |
-
-Cascade eval: **14b → 32b → Claude Opus** (102 cases × 3 seeds)
-
-| Metric | Result |
-|--------|--------|
-| Cascade accuracy | **100.0%** (mean, 3 seeds) |
-| Opus-solo etalon | 98.3% |
-| Δ vs Opus | **+1.7%** |
-| Traffic served by 14b | **99%** (101/102 cases avg) |
-| Traffic escalated to 32b | 1% (1/102 avg) — catches `save live state` → handoff edge case |
-| Traffic reaching Opus API | **0%** |
+## Training
 
-Fine-tuned cascade outperforms Claude Opus on `edge` (+16.7%) and `know` (+14.3%).
+- **Base model**: Qwen/Qwen3-32B (4-bit quantized for training)
+- **Method**: MLX LoRA SFT (rank=16, 8 layers, scale=20.0) × 14 iterative rounds
+- **Training data**: 300 eval-aligned prompts + targeted failure remediation per round
+- **Quantization**: Q4_K_M via llama.cpp (18 GB)
+- **Hardware**: Apple M5 Max 48 GB unified memory
 
-## Version History
+## Usage
 
-| Version | Base | BFCL | Notes |
-|---------|------|------|-------|
-| v7 (current) | Qwen3-30B-A3B MoE | **100.0% PERFECT** | Fixed: "what do I know + search memory" compound → knowledge_search |
-| v6 | Qwen3-30B-A3B MoE | 99.0% | Fixed MoE merge (BF16 safetensors + correct MLX→HF key mapping) |
-| v5 | Qwen3-30B-A3B MoE | 97.1% | 18× density fix; 9GB smaller, 4× faster vs dense |
-| v4 | Qwen3-30B-A3B MoE | 92.2% | rank=32 experiment — regressed vs v3 |
-| v3 | Qwen3-30B-A3B MoE | 92.5% | 20× reps + LR=1e-5 — hit rank bottleneck |
-| v2 | Qwen3-30B-A3B MoE | 92.5% | v34 corpus + 1400 iters |
-| v33 (dense) | Qwen3-32B dense | 99.0% | Prior generation — larger/slower |
+### Ollama
 
-## Tools
+```bash
+ollama pull dcostenco/prism-coder:32b
+ollama run dcostenco/prism-coder:32b "Load context for the billing-service project."
+```
 
-The model routes between exactly 6 tools:
+### llama.cpp
 
-1. `session_load_context` — load/fetch/resume project context
-2. `session_save_ledger` — note/log/remember/record progress
-3. `session_save_handoff` — handoff/relay to next agent/session
-4. `session_compact_ledger` — compact/archive/shrink ledger
-5. `session_search_memory` — recall past sessions/conversations
-6. `knowledge_search` — search stored notes/knowledge base
+```bash
+llama-cli -m prism-coder-32b-q4km.gguf \
+  -p "<|im_start|>system\nYou are Synalux...<|im_end|>\n<|im_start|>user\nLoad context for billing.<|im_end|>\n<|im_start|>assistant\n"
+```
 
-## Files
+## Model Family
 
-| File | Size | Use |
-|------|------|-----|
-| `qwen3-30b-a3b-v7-iq4nl.gguf` | 16 GB | **Current — recommended** |
-| `qwen3-30b-a3b-v6-iq4nl.gguf` | 17 GB | Previous (99.0%) |
-| `qwen3-30b-a3b-v5-iq4nl.gguf` | 17 GB | Previous (97.1%) |
-| `qwen3-32b-v33-q6k.gguf` | 25 GB | Dense predecessor (99.0%, legacy) |
+| Model | Size | eval_300 |
+|-------|------|----------|
+| prism-coder:1b7 | 2.2 GB | 100% |
+| prism-coder:4b | 2.5 GB | 100% |
+| prism-coder:14b | 9.0 GB | 99.7% |
+| **prism-coder:32b** | **18 GB** | **100%** |
 
-## Usage (Ollama)
+## License
 
-```bash
-ollama run dcostenco/prism-coder:32b
-```
+Apache 2.0
 
-## Training
+## Author
 
-- **Base**: Qwen/Qwen3-30B-A3B (HF BF16, ~57 GB)
-- **Adapters**: v6 LoRA (rank=8, scale=10, 8 layers, LR=1e-5)
-- **Merge**: Direct safetensors merge on HF BF16 base; delta = (scale/rank) × B^T A^T for attn/gate; delta[i] = (scale/rank) × B[i] A[i] for MoE experts (128 experts stacked)
-- **Key fix**: v5 merge used wrong base (MLX 4-bit, can't apply float LoRA delta) and uppercase regex `lora_[AB]` vs actual lowercase `lora_a`/`lora_b` adapter keys
-- **Hardware**: Apple Silicon (M-series, 64 GB RAM)
+[Synalux](https://synalux.com) — AI-powered clinical and development tools.

prism-coder-32b-q4km.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cd31306a03b67edfd9b7cb3f863f1c95e4fadee58f049bcbf7b9d4a9a953fed6
+size 19762149120