README.md

---
language: en
license: apache-2.0
tags:
  - tool-routing
  - function-calling
  - prism-coder
  - qwen3
  - gguf
  - synalux
base_model: Qwen/Qwen3-1.7B
---

# prism-coder:1b7 — 17-Tool Memory Agent (Always-Fits Tier)

Fine-tuned Qwen3-1.7B for full Prism Memory tool routing in the [Prism Coder](https://ollama.com/dcostenco/prism-coder) system.
Primary deployment: **any device** via llama.cpp GGUF — the ultra-lightweight tier.

## eval_300 Benchmark — swe43 (Current)

**300/300 × 3 shuffled runs = 100.0%, 0 flaky**

| Category | Count | Description | Accuracy |
|----------|------:|-------------|:--------:|
| natural_phrasing | 50 | Natural language → correct tool | 100% |
| adversarial_trap | 70 | Coding/CS questions → plain text (no tool) | 100% |
| disambiguation | 40 | Ambiguous session vs knowledge ops | 100% |
| edge_case | 25 | Self-description, capability queries → plain text | 100% |
| verifier | 25 | Verify-then-act chains | 100% |
| param_extraction | 25 | Extract project/query from prompt | 100% |
| cascade | 25 | Multi-step tool chains | 100% |
| multi_intent | 20 | Compound instructions | 100% |
| abstention | 20 | Greetings, math, creative requests → plain text | 100% |

300 test cases, 3 shuffled runs, temperature=0, 0 hallucinations across all runs.

## Tools

Routes to 17 Prism Memory tools + knows when NOT to call any tool:

| Tool | Trigger |
|------|---------|
| `session_load_context` | Load/resume project context, "starting fresh" |
| `session_save_ledger` | Log/record completed work |
| `session_save_handoff` | Create handoff note for next session |
| `session_search_memory` | Recall prior discussions |
| `session_forget_memory` | Delete a memory entry |
| `session_health_check` | Check session system health |
| `session_compact_ledger` | Compact/prune session ledger |
| `session_export_memory` | Export session data |
| `session_task_route` | Route task: local vs cloud |
| `session_save_experience` | Save a notable experience |
| `session_synthesize_edges` | Build session graph edges |
| `session_backfill_links` | Repair dangling session links |
| `knowledge_search` | Search stored knowledge base |
| `knowledge_forget` | Remove a knowledge entry |
| `knowledge_upvote` | Upvote knowledge entry |
| `knowledge_downvote` | Downvote knowledge entry |
| `knowledge_set_retention` | Set retention policy |

**Abstains (plain text)** for: coding questions, CS concepts, arithmetic, greetings, capability queries, creative requests, general knowledge.

## Version History

| Version | eval_300 | Notes |
|---------|---------|-------|
| swe43 | **300/300 × 3 runs = 100.0%** | Fresh rank=32 LoRA + `<think>` routing, Q8_0 GGUF |
| swe30 | 280/300 = 93.3% | Q8_0 first round (fixed Q4KM quantization erasure) |
| v43l | 203/300 = 67.7% | Baseline before SWE training |
| v42 | 100% BFCL 6-tool | Previous 6-tool routing model |

## Key Training Insights

- **Q8_0 quantization required** — Q4KM erased LoRA deltas for soft abstain patterns (87%→93% at R30)
- **Adapter saturation** — After 39 cumulative rounds at rank=8, adapter was saturated. Fresh rank=32 on R39-merged base broke plateau in one round (93.3%→99.7%)  
- **`<think>` routing blocks** — Added CoT reasoning to abstain examples activates Qwen3's pretrained thinking circuit, providing explicit gradient path for the routing decision

## Model Details

- **Base**: Qwen/Qwen3-1.7B → merged through 43 SWE training rounds
- **Format**: GGUF Q8_0 (2.2 GB)
- **Context**: 8,192 tokens
- **Final adapter**: MLX LoRA rank=32, all 28 layers, LR=3e-6→8e-7, 1,267 train rows/round
- **Total training**: 43 rounds of cumulative SFT + 4 fresh rank=32 rounds

## Usage

```bash
ollama pull dcostenco/prism-coder:1b7
ollama run dcostenco/prism-coder:1b7
```

Or via the [Synalux Prism MCP server](https://github.com/dcostenco/prism-mcp) which routes tool calls automatically.