---
license: apache-2.0
base_model: Qwen/Qwen2.5-7B-Instruct
tags:
  - web-agent
  - browser-agent
  - abliterated
  - function-calling
  - llama.cpp
  - gguf
---

# NanoWeb-7B-Instruct (Phase 2 - Multi-Domain Distillation)

NanoWeb-7B-Instruct v2 is an autonomous, sovereign, refusal-abliterated browser-agent model fine-tuned on **8,280 multi-domain web flight-recorder trajectories** (~31,000 tool calls) from base `Qwen/Qwen2.5-7B-Instruct`.

## Interaction Domains Covered
1. **E-Commerce & Price Intelligence**: Product navigation, price comparison, specification extraction, stock alerts.
2. **Academic & Technical Research**: ArXiv paper synthesis, GitHub repository inspection, documentation navigation.
3. **Financial & Market Intelligence**: Ticker analysis, earnings report lookups, crypto metrics, macroeconomic indicators.
4. **News & Geopolitical Tracking**: Breaking news cross-referencing, primary source verification, fact-checking.
5. **Travel, Hospitality & Flight Schedules**: Routing queries, hotel amenity comparisons, booking workflow traversal.
6. **Complex Multi-Hop Web Synthesis**: Multi-source reconciliation, adversarial error recovery, dynamic DOM element interaction.

## Tool Schemas
Pre-tuned with native Hermes function calling on:
- `fetch_page(url)`: Stealth headless retrieval with Jev-pruned numbered interactive DOM elements.
- `read_page(url, chunk_index)`: Sequential chunked reading for dense textual content.
- `find_in_page(url, pattern)`: Regex and keyword search across DOM elements.
- `browser_action(action, element_id, text, key, value)`: Interactive element execution (`click`, `type`, `select`, `press_enter`).
- `web_search(query)`: Multi-provider search querying.
- `finish_research(answer, sources)`: Final evidence synthesis.

## Mechanistic Refusal Abliteration
Refusal representations are mathematically annihilated via in-place orthogonal subspace projection ($W \leftarrow (I - \hat{r}\hat{r}^T)W$) across layers 14, 15, 16, 17, 18, 19, and 20 on both MLP down-projections and self-attention output projections ($P_\perp W_{\text{down}}$ & $P_\perp W_o$), eliminating false refusals on sovereign research, web scraping, and DOM automation.

## Quantized GGUF Deliverables (SHA-256 Verified)
- `nanoweb-7b-v2-instruct-Q4_K_M.gguf`: ~4.5 GB (Optimal for consumer GPUs / Ollama / llama.cpp)
- `nanoweb-7b-v2-instruct-Q8_0.gguf`: ~7.5 GB (High-precision near-FP16 fidelity)
- Accompanying `.sha256` manifest files for cryptographic integrity verification.