XcaliburBiz
A lightweight report writer. A compact data summary goes in and a professional report comes out.
What it is
- Base checkpoint:
Qwen/Qwen1.5-1.8B-Chat - Supervised pairs: 658
- Preference pairs: 1699
- Adapter rank: 16
License and attribution
This model is a derivative work based on Qwen1.5-1.8B-Chat by the Qwen Team, modified via Supervised Fine-Tuning and Direct Preference Optimization (DPO) by XcaliburMoon.
- License: Apache License 2.0 (https://www.apache.org/licenses/LICENSE-2.0), the license of the base checkpoint, carried forward unchanged.
- Publisher: XcaliburMoon
- Retained copies:
LICENSEandNOTICEin this repository. - ONNX conversion reference:
Xenova/Qwen1.5-1.8B-Chat.
Modifications
- Supervised fine-tuning of a low rank adapter over section pairs built from the house reports.
- Direct preference optimization from the supervised adapter.
- Merging the aligned adapter into the base weights at full precision.
- Quantization of the merged weights into the GGUF forms and one ONNX export.
Apache License 2.0 grants no right to the trade names of the licensor beyond the use that describes the origin of the work. The name Qwen and the names of the Qwen Team belong to their owners. The name appears in this repository to state where the base checkpoint came from, and it states no endorsement of this work.
House rules the model is trained to keep
- No em dash and no en dash.
- No hyphen with spaces around it used as a dash.
- No emoji.
- Every contraction written out in full.
- No worn prepositional phrase.
Measured on held-out material
| Form | Style clean | Figures kept | Passed | Seconds per report |
|---|---|---|---|---|
| base | 0.925 | 0.3372 | 0.225 | 10.826 |
| posttrain | 0.975 | 0.1244 | 0.075 | 4.697 |
| xcaliburbiz-Q2_K | 0.95 | 0.0798 | 0.075 | 0.485 |
| xcaliburbiz-Q3_K_M | 1.0 | 0.1084 | 0.1 | 0.38 |
| xcaliburbiz-Q4_K_M | 0.975 | 0.1011 | 0.075 | 0.683 |
| xcaliburbiz-Q5_K_M | 0.925 | 0.1423 | 0.125 | 1.004 |
| xcaliburbiz-Q6_K | 1.0 | 0.1203 | 0.05 | 1.002 |
| xcaliburbiz-Q8_0 | 0.975 | 0.1098 | 0.075 | 0.925 |
Files
LICENSE11357 bytesNOTICE1298 bytesadapter/README.md1506 bytesadapter/adapter_config.json1101 bytesadapter/adapter_model.safetensors12595704 bytesadapter/added_tokens.json80 bytesadapter/chat_template.jinja328 bytesadapter/merges.txt1671853 bytesadapter/special_tokens_map.json367 bytesadapter/tokenizer.json11418266 bytesadapter/tokenizer_config.json973 bytesadapter/vocab.json2776833 bytesadapter-sft/README.md1513 bytesadapter-sft/adapter_config.json1101 bytesadapter-sft/adapter_model.safetensors12595704 bytesadapter-sft/added_tokens.json80 bytesadapter-sft/chat_template.jinja328 bytesadapter-sft/merges.txt1671853 bytesadapter-sft/special_tokens_map.json367 bytesadapter-sft/tokenizer.json11418266 bytesadapter-sft/tokenizer_config.json973 bytesadapter-sft/vocab.json2776833 bytesmerged-bf16/README.md1566 bytesmerged-bf16/added_tokens.json80 bytesmerged-bf16/chat_template.jinja328 bytesmerged-bf16/config.json1230 bytesmerged-bf16/generation_config.json205 bytesmerged-bf16/merges.txt1671853 bytesmerged-bf16/model.safetensors3673690696 bytesmerged-bf16/special_tokens_map.json367 bytesmerged-bf16/tokenizer.json11418266 bytesmerged-bf16/tokenizer_config.json973 bytesmerged-bf16/vocab.json2776833 bytesgguf/quantize_record.json1101 bytesgguf/xcaliburbiz-Q2_K.gguf846570240 bytesgguf/xcaliburbiz-Q3_K_M.gguf1015880448 bytesgguf/xcaliburbiz-Q4_K_M.gguf1217753856 bytesgguf/xcaliburbiz-Q5_K_M.gguf1376776960 bytesgguf/xcaliburbiz-Q6_K.gguf1579027200 bytesgguf/xcaliburbiz-Q8_0.gguf1958304512 bytesbench.json284243 bytessft_record.json530 bytesdpo_record.json5586 bytesmerge_record.json187 bytesquantize_record.json1101 bytesdata/sft.jsonl2738739 bytesdata/dpo.jsonl8982824 bytesdata/eval.jsonl248519 bytesdata/manifest.json6204 bytes
The ONNX folder under gguf/ sits beside the quants for the in-browser engine.
Data
The pairs were built from the house reports. The supervised pairs carry a data summary in and a report section out; the preference pairs carry a good answer and a faulty one. The validation split is held out by report rather than by pair, so no measured answer was trained on.