---
base_model: mlx-community/Qwen3.6-35B-A3B-4bit
library_name: mlx
license: apache-2.0
language:
  - en
pipeline_tag: text-generation
tags:
  - lora
  - mlx
  - code
  - dependency-graph
  - static-analysis
  - qwen3-moe
  - oracle-verified
---

# depwire-slm

LoRA adapter for **Qwen3.6-35B-A3B** (MoE, 4-bit MLX), fine-tuned to reason over
codebase **dependency-graph facts** produced by the Depwire CLI — impact
analysis, security findings, dead code, blast radius — with every answer
deterministically verified against the same graph engine.

> **v0.4 — facts-in / reasoning-out, 100% grounded on the 1.16.0 oracle.**
> The model is *given* graph facts in the prompt (computed deterministically by
> depwire-cli). It reasons over them — prioritizing, sequencing, explaining risk —
> and never produces graph facts from its weights. Every claim is checked against
> the same depwire-cli engine via a deterministic oracle harness on a held-out set
> of 40 real repositories.
>
> **v0.3 (superseded)** reached 94.7% on the corrected 1.9.2 oracle. v0.4 is
> retrained on 9,918 pairs from depwire-cli 1.16.0 raw graph snapshots. The
> corrected oracle excludes fabricated `calls` edges from dependency structure;
> v0.4 reaches 100.0% grounding (130/130) with no format-level failures.
>
> **v0.1 (archived)** was a template-based proof of concept trained on
> generated pairs with a keyword-based benchmark. It is superseded.

## The architecture (why this is different)

```
depwire-cli (ground truth graph facts)
        │
        ▼
facts-in prompt ──► SLM reasons (priority, sequence, risk) ──► typed JSON answer
        │                                                       │
        └──────────────────► oracle_eval.py verifies claims ────┘
```

- **Facts go in the prompt, never invented by the model.** If a file, symbol,
  or number is not in the prompt's FACTS block, the model must not claim it.
- **Every claim is oracle-checked.** `claims` fields are compared
  deterministically against live `depwire` output (exact enums, exact counts,
  subset-of-truth file lists).
- **The reasoning fields are free text** — judged by quality, never able to
  cause a grounding failure.

## Results (v0.4) — oracle grounding on held-out repos

130 evaluation cases across 40 held-out repositories (disjoint from all
training repos). Every case: prompt contains real depwire facts; model answer
checked against live depwire output (depwire-cli 1.16.0 oracle). Import edges
alone define dependency structure; `calls` edges are excluded.

| Task       | Grounding  | Cases |
|------------|-----------|-------|
| Impact     | **100.0%** | 63/63 |
| Security   | **100.0%** | 30/30 |
| Dead code  | **100.0%** | 37/37 |
| **Overall**| **100.0%** | **130/130** |

- Critical hallucinations: **0**
- Phantom vulnerabilities: **0**
- Oracle errors (excluded): **0**
- Schema/format failures: **0**
- Best checkpoint: iter 4200, val loss 0.091
- The dead-code confidence casing, security severity casing, and token-limit
  failures observed in earlier versions are absent in this evaluation

Baseline context: the same oracle harness measured the v0.1 template model at
**1.0% grounding** on held-out impact cases — it could not produce verifiable
answers about code it wasn't given facts for. v0.4's 100.0% measures a
different, deliberately narrower capability: *given the facts, reason over
them without contradicting the oracle.* The score applies to this fixed
130-case held-out set; it is not a claim of perfect behavior on arbitrary
repositories or prompts.

### Version history

| Version | Oracle | Evaluation |
|---------|--------|------------|
| v0.1 | depwire-cli 1.8.7 | 55% keyword benchmark |
| v0.2 | depwire-cli 1.8.7 | 95.4% oracle grounding |
| v0.2 | depwire-cli 1.9.2 | 93.1% on corrected parser |
| v0.3 | depwire-cli 1.9.2 | 94.7% oracle grounding |
| **v0.4** | **depwire-cli 1.16.0** | **100.0% (130/130)** |

## Usage (MLX on Apple Silicon)

**Important: this is a facts-in model. You must provide the graph facts in
the prompt.** It will not (and should not) invent them.

```bash
pip install mlx-lm
mkdir -p lora-adapters
huggingface-cli download atefataya/depwire-slm adapters.safetensors \
  adapter_config.json \
  --local-dir lora-adapters

mlx_lm.generate \
  --model mlx-community/Qwen3.6-35B-A3B-4bit \
  --adapter-path lora-adapters \
  --prompt "FACTS (from depwire):
Target file: src/auth/service.ts
Total affected: 3
Affected: src/api/users.ts, src/api/admin.ts, src/api/auth.ts

Task: Which affected files should be checked first and why? Answer as JSON."
```

For production use, the Depwire CLI computes the facts; this adapter
reasons over them. The eval harness (`scripts/oracle_eval.py`) is Apache-2.0
and calls depwire-cli as an external subprocess.

## Training Data

9,918 oracle-passing teacher-distilled pairs across three tasks, generated
from real depwire-cli 1.16.0 output on non-held-out repositories. The raw graph
schema uses countable symbols and import-only dependency edges, removing
fabricated call-edge relationships from the training facts:

- Impact: affected-files analysis from depwire facts
- Dead code: dead-symbol analysis from depwire facts
- Security: vulnerability findings from depwire facts

Every training completion passed the same oracle harness used for evaluation
(teacher = DeepSeek; outputs filtered through `oracle_eval.py`; contradictions
discarded). Training data is strictly disjoint from the 40 evaluation repos.

| Metric              | Value |
|---------------------|-------|
| Oracle-passing pairs generated | 9,918 |
| Training split (mlx/train.jsonl) | 9,422 |
| Validation split (mlx/valid.jsonl) | 496 |
| LoRA                | rank 8, 16 layers, scale 20.0 |
| Base                | Qwen3.6-35B-A3B-4bit (MoE, 4-bit) |
| Max seq             | 1,024 tokens |
| Loss                | val 0.091 (iter 4200) |

## Limitations

- **Facts-in only.** The model produces graph facts from its weights at
  ~0% accuracy — by design. Always feed it the FACTS block; never ask it to
  recall a repository's structure from memory.
- **Reasoning ceiling = teacher's.** Distillation means the model's reasoning
  quality is bounded by the teacher model used to generate completions.
- **It does not read code.** It reasons about graph numbers you provide
  (dependents, in-degree, severity, dead-code confidence).
- **Non-English prompts are out of scope.**
- **Java, Ruby, C++, Swift, Kotlin** graphs were excluded from training (tree-sitter
  WASM parser crashes); the adapter has not seen those graphs.
- **Evaluation scope.** The 100% result is on a fixed 130-case held-out set with
  facts supplied in the prompt and deterministic oracle grading. Broader prompt
  styles, repositories, and reasoning quality require separate evaluation.

## Repository

Pipeline, oracle harness, and eval are open source:
https://github.com/atef-ataya/depwire-slm (tag `v0.4`)

## License

Apache 2.0. Base model (`mlx-community/Qwen3.6-35B-A3B-4bit`) is subject to its
own license — check the Qwen model card before use.