# qwen3-1.7b-ben-franklin-thinking-v2-lora

## Summary

Custom Benjamin Franklin LoRA adapter for Qwen3 1.7B 4-bit.

## Paths

- Portfolio copy: `adapters/qwen3-1.7b-ben-franklin-thinking-v2-lora`
- Source: copied adapter artifact in `adapters/qwen3-1.7b-ben-franklin-thinking-v2-lora`
- Base model: `unsloth/Qwen3-1.7B-unsloth-bnb-4bit`

## Adapter details

- Adapter directory size: 0.78 GB
- adapter_model.safetensors: 133.0 MB
- SHA256(adapter_model.safetensors): `0a1e6a1f82eac2f68860514f3ec98ef9794f3b4f7410798efdf9e07dbffe8f6a`
- LoRA rank: 32
- LoRA alpha: 64
- Target modules: `['q_proj', 'v_proj', 'down_proj', 'k_proj', 'gate_proj', 'o_proj', 'up_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.

## 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.

## Data mix

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

## Performance

Not benchmarked in the current coherence index.

## 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.