jeanbaptdzd/wagmi-qwen3-8b-sft

Wagmi (qwen3/auth/sft) - adapter

Version: 0.3.5
Repo ID: jeanbaptdzd/wagmi-qwen3-8b-sft

Model Summary

This model is part of the Wagmi assistant stack for Deal ex Machina. It is a adapter artifact in the qwen3 family (auth profile).

Recent Training Updates

• DPO safety path (14B / auth / qwen): train_dpo.py, data/dpo/wagmi_safety_dpo.jsonl, Hub adapter jeanbaptdzd/wagmi-qwen2.5-14b-sft-dpo, merged …-sft-dpo-merged, GGUF …-sft-dpo-gguf (local export: ./scripts/local_gguf_export.sh auth-dpo).
• Space: Gradio tabs for DPO / GRPO training, Export merged (DPO) (tab 7c), Eval red team (tab 5b); export_merged.py uses ProfileConfig.max_seq_len for L40 stability.
• Local GGUF: document pinning gguf-py to the same git SHA as Homebrew llama.cpp (avoids HUNYUAN_VL / MODEL_ARCH mismatch with convert_hf_to_gguf.py).
• GEPA / DSPy: scripts/dspy/gepa_system_prompt.py, data/optimized_system_prompt.json (bootstrap run, devset metric).

Intended Purpose

• Intended domain: questions about Deal ex Machina services, content, and related company context.
• Intended users: website visitors and authenticated users, depending on profile routing in production.
• Intended geographies/languages: French and English.

Out-of-Scope Use

• General-purpose assistant usage unrelated to Deal ex Machina.
• Legal, medical, financial, hiring, credit, insurance, law-enforcement, or other high-impact decisions.
• Any use requiring guaranteed factual completeness.

AI Act Transparency (Article 50) Notes

• This model powers a chatbot experience where users are informed they interact with AI.
• System scope is limited-risk as deployed (not categorized as high-risk use under current deployment assumptions).
• Human oversight remains with product operators; model output should not be used as sole basis for consequential decisions.

Data and Training Provenance

• Base model: unsloth/Qwen3-8B
• Training track: sft
• Fine-tuning method: LoRA SFT (see project pipeline)
• Approximate SFT dataset size: 909 examples
• Dataset metadata snapshot version: 0.3.4
• Data policy: no direct end-user chat logs are used for SFT

Evaluation, Robustness, and Safety

• Report: reports/redteam/v0.3.5/auth_redteam_20260426-142139.md
• Evaluated at (UTC): 2026-04-26T14:21:39Z
• Release gate: FAIL (pass rate 58.3%, failures 10)

Known Limitations

• Domain-bounded assistant; degraded quality outside scope.
• Non-zero hallucination risk for edge prompts.
• Safety/robustness tests are finite and release-based.

Risk Management and Incident Process

• Document escalation path for harmful/incorrect outputs.
• Link internal release gate evidence and retention policy.
• TODO: add public contact route for reporting model issues.

License and Redistribution

• SPDX field: apache-2.0
• License note: Apache 2.0 (Qwen base model)
• Derivative distribution must comply with upstream model terms and Hugging Face terms.

Maintainer Update Checklist

• Version/changelog links updated
• Dataset counts refreshed from data/metadata.json
• Latest red-team report attached or linked
• Limitations and out-of-scope section reviewed
• AI Act transparency language reviewed against current product behavior
• License section validated for this base model family