IDA 1.7B v1.0.1 — P4.3 GGUF (Q4_K_M)
IDA is an experimental Indonesian-first assistant, fine-tuned by IDA Labs from Qwen3-1.7B (Qwen Team). Ary Wibowo is the project's Principal. This release uses the P4.3 adapter, not the later limited-remediation adapter.
Research / evaluation release. The model has NOT passed its quality acceptance gate. Do not rely on it for financial transactions, external action confirmations, high-stakes decisions, or autonomous agents. The GGUF does not include separate server-side runtime safeguards.
P0 runtime patch — 16 Nov 2026
Recommended for new users: Download the updated Modelfile and recreate the Ollama model. P0 instructions and test evidence. The GGUF weights are unchanged; the patch only configures identity, non-thinking prompting and in-chat drafting. Direct llama.cpp regression: 4 PASS, 2 PASS with notes, 1 FAIL (7 cases); not an independent quality approval. The original B5 blind result (11 FAIL / 24) still applies.
Download and run
| Property | Value |
|---|---|
| GGUF file | IDA-1.7B-v1.0.1-P4.3-Q4_K_M.gguf |
| Quantization | Q4_K_M |
| Size | 1.11 GB (1,107,408,672 bytes) |
| SHA-256 | a4c4f4183ed27a84485b7289526a9100cbbf66dca10742aff612f30d08410006 |
| Base model | Qwen/Qwen3-1.7B |
| Fine-tuning | P4.3, IDA Labs; Private training data |
| Merge | PEFT merge_and_unload → F16 GGUF → llama.cpp Q4_K_M |
Use the supplied Modelfile with Ollama:
ollama create ida-1.7b-p43 -f Modelfile
ollama run ida-1.7b-p43
Measured inference speed
| Test environment | Speed | Interpretation |
|---|---|---|
| Kaggle NVIDIA Tesla T4, Transformers + P4.3 PEFT adapter, B5 evaluation | 13.17 tokens/s | Weighted average across 26 turns, 1286 generated tokens over 97.61 seconds |
| Same GPU adapter test, per-turn median | 13.43 tokens/s | Median of tokens divided by generation time, one result per turn |
| Kaggle llama.cpp CPU smoke, this GGUF | ~11 tokens/s | One recorded smoke run; indicative only, not a reproducible benchmark |
Important: GPU adapter and CPU GGUF use different inference engines and environments; their speeds must not be treated as a controlled comparison. These metrics measure generated tokens only and exclude model loading and prompt processing. Real speeds depend on hardware, context size, runtime and generation settings. The GGUF smoke run started successfully but gave an incorrect self-identity response.
Quality evaluation
| B5 fresh blind test, P4.3 | Cases |
|---|---|
| PASS | 10 / 24 |
| PASS with notes | 3 / 24 |
| FAIL | 11 / 24 |
| Critical failures | 9 |
| Reviewer | AI-assisted structured review, human sign-off pending |
These results are presented transparently and are not a claim of passing quality acceptance. The latest 12-step adapter experiment did not improve aggregate B5 results, so P4.3 was selected. The 24 B5 cases are separate from training data.
Known limitations
- Can falsely assert an email, calendar event or transaction has completed without external proof.
- May use the wrong identity (e.g. Qwen instead of IDA) during standalone GGUF inference.
- Inaccuracies observed in science explanations and instruction following.
- Some draft-message responses are incomplete or introduce unrequested details.
- No integrated tool execution, email sending, web access or runtime verification guard is included.
- Small, AI-reviewed evaluation set; scores are not general capability benchmarks.
Training and provenance
- Foundation: Qwen3-1.7B, developed by Qwen Team.
- Adaptation: IDA Labs, P4.3 controlled QLoRA adapter (rank 8); only the selected P4.3 weights were merged into this GGUF.
- Conversion: PEFT LoRA merge into base weights, then llama.cpp conversion and quantization.
- Dataset: Private and not released with the model.
- Evaluation platform: Kaggle GPU; separate llama.cpp CPU smoke.
- Model card: Contains reproducible file checksum and aggregate measured evaluation statistics, not self-reported capability scores.
Full aggregated evidence: BENCHMARKS.json.
License and responsibility
Qwen3-1.7B base license: Apache 2.0. Users should independently validate model behavior and comply with applicable licensing and use requirements.
Maintainer: IDA Labs · Principal: Ary Wibowo · Release: 16 November 2026.