Instructions to use dcostenco/prism-coder-1.7b with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use dcostenco/prism-coder-1.7b with llama-cpp-python:

# !pip install llama-cpp-python

from llama_cpp import Llama

llm = Llama.from_pretrained(
	repo_id="dcostenco/prism-coder-1.7b",
	filename="prism-aac-1b7-q4km.gguf",
)

llm.create_chat_completion(
	messages = "No input example has been defined for this model task."
)

Notebooks
Google Colab
Kaggle
Local Apps Settings

llama.cpp

How to use dcostenco/prism-coder-1.7b with llama.cpp:

Install from brew

brew install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama-server -hf dcostenco/prism-coder-1.7b:Q8_0
# Run inference directly in the terminal:
llama-cli -hf dcostenco/prism-coder-1.7b:Q8_0

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama-server -hf dcostenco/prism-coder-1.7b:Q8_0
# Run inference directly in the terminal:
llama-cli -hf dcostenco/prism-coder-1.7b:Q8_0

Use pre-built binary

# Download pre-built binary from:
# https://github.com/ggerganov/llama.cpp/releases
# Start a local OpenAI-compatible server with a web UI:
./llama-server -hf dcostenco/prism-coder-1.7b:Q8_0
# Run inference directly in the terminal:
./llama-cli -hf dcostenco/prism-coder-1.7b:Q8_0

Build from source code

git clone https://github.com/ggerganov/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-server llama-cli
# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama-server -hf dcostenco/prism-coder-1.7b:Q8_0
# Run inference directly in the terminal:
./build/bin/llama-cli -hf dcostenco/prism-coder-1.7b:Q8_0

Use Docker

docker model run hf.co/dcostenco/prism-coder-1.7b:Q8_0

LM Studio
Jan
Ollama
How to use dcostenco/prism-coder-1.7b with Ollama:
```
ollama run hf.co/dcostenco/prism-coder-1.7b:Q8_0
```

Unsloth Studio

How to use dcostenco/prism-coder-1.7b with Unsloth Studio:

Install Unsloth Studio (macOS, Linux, WSL)

curl -fsSL https://unsloth.ai/install.sh | sh
# Run unsloth studio
unsloth studio -H 0.0.0.0 -p 8888
# Then open http://localhost:8888 in your browser
# Search for dcostenco/prism-coder-1.7b to start chatting

Install Unsloth Studio (Windows)

irm https://unsloth.ai/install.ps1 | iex
# Run unsloth studio
unsloth studio -H 0.0.0.0 -p 8888
# Then open http://localhost:8888 in your browser
# Search for dcostenco/prism-coder-1.7b to start chatting

Using HuggingFace Spaces for Unsloth

# No setup required
# Open https://huggingface.co/spaces/unsloth/studio in your browser
# Search for dcostenco/prism-coder-1.7b to start chatting

How to use dcostenco/prism-coder-1.7b with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama-server -hf dcostenco/prism-coder-1.7b:Q8_0

Configure the model in Pi

# Install Pi:
npm install -g @mariozechner/pi-coding-agent
# Add to ~/.pi/agent/models.json:
{
  "providers": {
    "llama-cpp": {
      "baseUrl": "http://localhost:8080/v1",
      "api": "openai-completions",
      "apiKey": "none",
      "models": [
        {
          "id": "dcostenco/prism-coder-1.7b:Q8_0"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Hermes Agent new

How to use dcostenco/prism-coder-1.7b with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama-server -hf dcostenco/prism-coder-1.7b:Q8_0

Configure Hermes

# Install Hermes:
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
hermes setup
# Point Hermes at the local server:
hermes config set model.provider custom
hermes config set model.base_url http://127.0.0.1:8080/v1
hermes config set model.default dcostenco/prism-coder-1.7b:Q8_0

Run Hermes

hermes

Docker Model Runner
How to use dcostenco/prism-coder-1.7b with Docker Model Runner:
```
docker model run hf.co/dcostenco/prism-coder-1.7b:Q8_0
```

Lemonade

How to use dcostenco/prism-coder-1.7b with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull dcostenco/prism-coder-1.7b:Q8_0

Run and chat with the model

lemonade run user.prism-coder-1.7b-Q8_0

List all available models

lemonade list

dcostenco commited on 13 days ago

Commit

625b3be

verified ·

1 Parent(s): 17886ae

Upload README.md with huggingface_hub

Browse files

Files changed (1) hide show

README.md +59 -45

README.md CHANGED Viewed

@@ -4,76 +4,90 @@ license: apache-2.0
 tags:
   - tool-routing
   - function-calling
-  - prism-aac
   - qwen3
   - gguf
 base_model: Qwen/Qwen3-1.7B
 ---
-# prism-coder:1.7b — Tool Routing Model (Always-Fits Tier)
-Fine-tuned Qwen3-1.7B for 6-tool routing in the [Prism AAC](https://github.com/dcostenco/prism-aac) system.
-Primary deployment: **any iOS device** via llama.cpp GGUF — the guaranteed fallback for all device tiers.
-## BFCL Routing Benchmark — v42 (Current)
-**Mean: 100.0%** (3-seed average, seeds 2027/2028/2029, 102 cases each)
 | 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% |
-Eval: MLX inference + thinking, temperature=0, 3-seed mean.
-Gate: ≥90% = deploy.
-## Version History
-| Version | BFCL | Notes |
-|---------|------|-------|
-| v42 | **100.0%** | Fixed 4 deterministic failures: cmpct tool name, compound edge, write-code irrel, pull-context load |
-| v41 | 96.1% | Proper safetensors merge — fixes mlx_lm.fuse LoRA loss |
-| v36 | 94.1% | LoRA rank=16, all 28 layers, mask-prompt |
-| v19 | ~88% | Baseline 1.7B routing |
 ## Tools
-The model routes to exactly 6 tools:
 | Tool | Trigger |
 |------|---------|
-| `session_load_context` | Load/resume/pull project context |
-| `session_save_ledger` | Note/log/record/remember something |
-| `session_save_handoff` | Pass state to next agent/session |
-| `session_compact_ledger` | Compact/shrink/prune ledger |
-| `session_search_memory` | Recall prior session discussions |
-| `knowledge_search` | Search stored knowledge base ("what do I know") |
-Plain text (no tool) for: AAC phrases, translations, weather, general facts, code/regex/functions, math.
 ## Model Details
-- **Base**: Qwen/Qwen3-1.7B
-- **Format**: GGUF Q4_K_M (~1.2 GB)
-- **Context**: 32,768 tokens
-- **Training**: MLX LoRA, rank=16, all 28 layers, 800 iters, LR=5e-5, v42 corpus (1028 train / 79 valid)
-- **Merge**: direct safetensors merge (scale/rank × B.T @ A.T) → llama.cpp convert → Q4_K_M quantization
 ## Usage
 ```bash
 ollama pull dcostenco/prism-coder:1b7
-ollama run prism-coder:1b7
 ```
-Or in [Prism AAC](https://github.com/dcostenco/prism-aac) — the app downloads and loads this model automatically on devices with <8 GB RAM.

 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.