# qwen2.5-7b-ben-franklin-v3-factual-coherence-r4-qv

## Summary

Custom Benjamin Franklin LoRA adapter for Qwen2.5 7B Instruct 4-bit.

## Paths

- Portfolio copy: `adapters/qwen2.5-7b-ben-franklin-v3-factual-coherence-r4-qv`
- Source: copied adapter artifact in `adapters/qwen2.5-7b-ben-franklin-v3-factual-coherence-r4-qv`
- Base model: `unsloth/Qwen2.5-7B-Instruct-bnb-4bit`

## Adapter details

- Adapter directory size: 0.06 GB
- adapter_model.safetensors: 4.8 MB
- SHA256(adapter_model.safetensors): `f76a2640a84889c7370e8cf8c4c0c6cc71e53a546a8cf7008f8c90ee7588769b`
- LoRA rank: 4
- LoRA alpha: 8
- Target modules: `['q_proj', 'v_proj']`

## Strengths

- Largest Benjamin Franklin LoRA family proven trainable on this RTX 3070 8GB machine.
- Best base reasoning/coherence among the local Franklin adapters.
- Good short-turn continuity in the coherence benchmark.
- Best numeric coherence benchmark score among evaluated 7B adapters.
- Benchmark score: 35 with flags {'base_identity_leak': 1}.

## Weaknesses / caveats

- Qwen2.5 7B base is already highly steerable, so improvements over the prompted base are modest.
- q/v-only LoRA is weak for implanting stubborn factual corrections.
- Craven Street/Hewson factuality remains unreliable unless retrieval/prompt context is supplied.
- Despite the name, not a clean factual fix: Craven Street answer still hallucinated.

## Data mix

- franklin_qwen3_8b_chatml_answer_clean.jsonl: 2400 rows
- franklin_7b_coherence_repair_v2.jsonl: 287 rows
- franklin_7b_factual_coherence_repair_v3.jsonl: 308 rows

## Performance

Score: 35

Flags: `{'base_identity_leak': 1}`

Raw benchmark: benchmarks/franklin_coherence/franklin_coherence_20260623_082910.json

HTML: benchmarks/franklin_coherence/franklin_coherence_20260623_082910.html

## Memory / compute requirements

Inference: RTX 3070 8GB works in 4-bit with the adapter; expect roughly 6-7GB VRAM. Training proven only with q_proj/v_proj, r=4 or r=8, max_seq_length=512, batch_size=1, gradient_accumulation=16; full-module LoRA is not recommended on 8GB.

## Possible project uses

- Conversational Franklin persona in local apps.
- A strong baseline for RAG-backed historical character demos.
- Best starting point for further 7B experiments.