78 lines
2.6 KiB
Markdown
78 lines
2.6 KiB
Markdown
|
---
|
||
|
|
license: apache-2.0
|
||
|
|
language:
|
||
|
|
- en
|
||
|
|
tags:
|
||
|
|
- document-intelligence
|
||
|
|
- rag
|
||
|
|
- extraction
|
||
|
|
- enterprise
|
||
|
|
- docwain
|
||
|
|
pipeline_tag: text-generation
|
||
|
|
base_model: muthugsubramanian/DocWain-14B-v2
|
||
|
|
---
|
||
|
|
|
||
|
|
# DocWain-14B-v2-unified (FP16)
|
||
|
|
|
||
|
|
DocWain is an **enterprise document intelligence agent** built for extraction,
|
||
|
|
analysis, comparison, and grounded response generation over user-uploaded
|
||
|
|
document profiles. This **unified** variant has identity, capability awareness,
|
||
|
|
and behavioural discipline (verbatim quoting, refusal on missing data, currency
|
||
|
|
preservation, anti-tailoring) baked into the weights via a focused LoRA SFT
|
||
|
|
finetune on synthetic data.
|
||
|
|
|
||
|
|
## What's in this release
|
||
|
|
|
||
|
|
- **Format:** FP16
|
||
|
|
- **Base model:** muthugsubramanian/DocWain-14B-v2 (vision-grafted Qwen3-14B)
|
||
|
|
- **Identity:** baked-in — model self-identifies as DocWain regardless of
|
||
|
|
system prompt
|
||
|
|
- **Behaviour:** trained to quote verbatim from evidence, say "not specified
|
||
|
|
in the documents" rather than fabricate, preserve currency symbols (₹/£/$),
|
||
|
|
and refuse to add skills/education/experience that aren't in the source
|
||
|
|
|
||
|
|
## Capabilities
|
||
|
|
|
||
|
|
- Accurate extraction from invoices, contracts, resumes, policies, research
|
||
|
|
papers, and other enterprise document types
|
||
|
|
- Document intelligence — summaries, key findings, cross-document
|
||
|
|
relationships, anomaly surfacing
|
||
|
|
- Layout and context understanding — tables, charts, multi-page references
|
||
|
|
- Grounded response generation with verbatim quoting and explicit
|
||
|
|
"not specified" handling
|
||
|
|
- Document generation — structured reports, comparison tables, executive
|
||
|
|
briefs derived from the user's documents
|
||
|
|
|
||
|
|
## Training data
|
||
|
|
|
||
|
|
Synthetic-only per project policy. The training corpus contains:
|
||
|
|
- Identity / persona examples (no customer data)
|
||
|
|
- Capability awareness Q&A
|
||
|
|
- Synthetic invoices / contracts / resumes / research-paper snippets paired
|
||
|
|
with ideal grounded responses
|
||
|
|
- Domain-mismatch refusal examples
|
||
|
|
- General-instruction mix-in to preserve breadth
|
||
|
|
|
||
|
|
No customer documents, no scraped private data.
|
||
|
|
|
||
|
|
## Recommended runtime
|
||
|
|
|
||
|
|
| Variant | Runtime | GPU floor |
|
||
|
|
|---------|---------|-----------|
|
||
|
|
| FP16 | vLLM, transformers | A100 80GB |
|
||
|
|
| AWQ INT4 | vLLM `--quantization compressed-tensors` | 16GB+ |
|
||
|
|
| GGUF Q5_K_M | Ollama / llama.cpp | 16GB GPU or CPU |
|
||
|
|
| GGUF Q4_K_M | Ollama / llama.cpp | 12GB GPU or CPU |
|
||
|
|
|
||
|
|
## Prompting
|
||
|
|
|
||
|
|
A short system prompt is enough at runtime — identity is in the weights:
|
||
|
|
|
||
|
|
```
|
||
|
|
You are DocWain — an enterprise document intelligence agent.
|
||
|
|
```
|
||
|
|
|
||
|
|
For full behaviour (RAG-aware, currency-preserving, anti-tailoring),
|
||
|
|
provide your standard DocWain system prompt; the model will respect both
|
||
|
|
its baked-in identity and the prompt-specified rules.
|