| <!doctype html><html><head><meta charset='utf-8'><title>Qwen (Ben) Franklin Model Zoo</title> |
| <style> |
| :root{--bg:#101317;--panel:#181d24;--text:#eef2f6;--muted:#aeb8c4;--line:#303846;--accent:#f2c166;--good:#9fe6a0;--bad:#ff9d9d} |
| body{margin:0;font-family:Inter,system-ui,Segoe UI,Arial,sans-serif;background:linear-gradient(135deg,#0d1117,#18202b);color:var(--text)} |
| header{padding:56px 7vw;border-bottom:1px solid var(--line);background:radial-gradient(circle at 20% 10%,rgba(242,193,102,.18),transparent 32%)} |
| h1{font-size:52px;margin:0 0 8px} h2{margin-top:38px} a{color:#9fd0ff} .lead{font-size:20px;max-width:980px;color:var(--muted);line-height:1.55} |
| main{padding:28px 7vw 70px} .panel,.model-card{background:rgba(24,29,36,.92);border:1px solid var(--line);border-radius:16px;padding:22px;margin:18px 0;box-shadow:0 8px 30px rgba(0,0,0,.25)} |
| table{border-collapse:collapse;width:100%;font-size:14px} th,td{border-bottom:1px solid var(--line);padding:10px;text-align:left;vertical-align:top} th{color:var(--accent)} |
| .cols{display:grid;grid-template-columns:1fr 1fr;gap:18px} code{display:block;white-space:pre-wrap;color:#d6e7ff;background:#0c1015;border:1px solid #273140;border-radius:8px;padding:10px;overflow:auto} |
| .badge{display:inline-block;padding:4px 8px;border:1px solid var(--line);border-radius:999px;color:var(--muted);margin:3px} |
| input{width:100%;box-sizing:border-box;padding:14px;border-radius:10px;border:1px solid var(--line);background:#0b1016;color:var(--text);font-size:16px} |
| @media(max-width:900px){.cols{grid-template-columns:1fr}h1{font-size:38px}} |
|
|
| </style></head><body> |
| <header><h1>Qwen (Ben) Franklin</h1><p class='lead'>A local model zoo of custom Benjamin Franklin LoRA adapters trained on Qwen-family bases. These models explore a living, useful Franklin voice: printerly wit, civic practicality, identity persistence, English-only dialogue, tool-call cleanup, factual biography repair, and more coherent conversation.</p> |
| <p><span class='badge'>22 custom Franklin adapters copied</span><span class='badge'>23.96 GB adapter artifacts</span><span class='badge'>RTX 3070 8GB tested</span></p></header> |
| <main> |
| <section class='panel'><h2>About</h2><p>“Qwen (Ben) Franklin” is a set of local LoRA experiments that turn compact Qwen-family models into a Benjamin Franklin conversational persona. The adapters range from fast 1.7B prototypes through 4B cleanup/factual variants to the newer 7B Qwen2.5 experiments, which are the largest Franklin LoRAs proven trainable on this machine.</p> |
| <p>The key lesson: the 7B models feel more coherent, but stubborn factual corrections such as the Craven Street bones story still benefit from retrieval or prompt-context. Some 4B variants contain stronger factual phrasing but suffer offline tool_call-tag regressions.</p></section> |
| <section class='panel'><h2>Performance table</h2><table><tr><th>Model</th><th>Base family</th><th>GB</th><th>r</th><th>Score</th><th>Flags</th><th>Card</th></tr><tr><td>qwen2.5-7b-ben-franklin-v1-lite-r4-qv</td><td>Qwen2.5 7B Instruct 4-bit</td><td>0.06</td><td>4</td><td>30</td><td>{'base_identity_leak': 1, 'overdenial': 1}</td><td><a href='model_cards/qwen2.5-7b-ben-franklin-v1-lite-r4-qv.md'>card</a></td></tr><tr><td>qwen2.5-7b-ben-franklin-v2-coherence-r4-qv</td><td>Qwen2.5 7B Instruct 4-bit</td><td>0.06</td><td>4</td><td>33</td><td>{'base_identity_leak': 1}</td><td><a href='model_cards/qwen2.5-7b-ben-franklin-v2-coherence-r4-qv.md'>card</a></td></tr><tr><td>qwen2.5-7b-ben-franklin-v3-factual-coherence-r4-qv</td><td>Qwen2.5 7B Instruct 4-bit</td><td>0.06</td><td>4</td><td>35</td><td>{'base_identity_leak': 1}</td><td><a href='model_cards/qwen2.5-7b-ben-franklin-v3-factual-coherence-r4-qv.md'>card</a></td></tr><tr><td>qwen2.5-7b-ben-franklin-v3c-factual-r8-qv-from-base</td><td>Qwen2.5 7B Instruct 4-bit</td><td>0.08</td><td>8</td><td>28</td><td>{'base_identity_leak': 1, 'overdenial': 1}</td><td><a href='model_cards/qwen2.5-7b-ben-franklin-v3c-factual-r8-qv-from-base.md'>card</a></td></tr><tr><td>qwen3-4b-instruct-2507-ben-franklin-v1-lora</td><td>Qwen3 4B Instruct 4-bit</td><td>1.34</td><td>16</td><td></td><td></td><td><a href='model_cards/qwen3-4b-instruct-2507-ben-franklin-v1-lora.md'>card</a></td></tr><tr><td>qwen3-4b-instruct-2507-ben-franklin-v2-chatml-lora</td><td>Qwen3 4B Instruct 4-bit</td><td>2.19</td><td>32</td><td></td><td></td><td><a href='model_cards/qwen3-4b-instruct-2507-ben-franklin-v2-chatml-lora.md'>card</a></td></tr><tr><td>qwen3-4b-instruct-2507-ben-franklin-v3-chatml-completions-lora</td><td>Qwen3 4B Instruct 4-bit</td><td>2.19</td><td>32</td><td></td><td></td><td><a href='model_cards/qwen3-4b-instruct-2507-ben-franklin-v3-chatml-completions-lora.md'>card</a></td></tr><tr><td>qwen3-4b-instruct-2507-ben-franklin-v4-toolcall-clean-lora</td><td>Qwen3 4B Instruct 4-bit</td><td>1.81</td><td>32</td><td></td><td></td><td><a href='model_cards/qwen3-4b-instruct-2507-ben-franklin-v4-toolcall-clean-lora.md'>card</a></td></tr><tr><td>qwen3-4b-instruct-2507-ben-franklin-v5-english-lock-lora</td><td>Qwen3 4B Instruct 4-bit</td><td>1.81</td><td>32</td><td>-92</td><td>{'tool_call': 15, 'continuity_miss': 1}</td><td><a href='model_cards/qwen3-4b-instruct-2507-ben-franklin-v5-english-lock-lora.md'>card</a></td></tr><tr><td>qwen3-1.7b-ben-franklin-identity-reinforced-lora</td><td>Qwen3 1.7B 4-bit</td><td>0.99</td><td>32</td><td></td><td></td><td><a href='model_cards/qwen3-1.7b-ben-franklin-identity-reinforced-lora.md'>card</a></td></tr><tr><td>qwen3-1.7b-ben-franklin-openai-expanded-lora</td><td>Qwen3 1.7B 4-bit</td><td>1.2</td><td>32</td><td></td><td></td><td><a href='model_cards/qwen3-1.7b-ben-franklin-openai-expanded-lora.md'>card</a></td></tr><tr><td>qwen3-1.7b-ben-franklin-thinking-lora</td><td>Qwen3 1.7B 4-bit</td><td>0.99</td><td>32</td><td></td><td></td><td><a href='model_cards/qwen3-1.7b-ben-franklin-thinking-lora.md'>card</a></td></tr><tr><td>qwen3-1.7b-ben-franklin-thinking-v2-lora</td><td>Qwen3 1.7B 4-bit</td><td>0.78</td><td>32</td><td></td><td></td><td><a href='model_cards/qwen3-1.7b-ben-franklin-thinking-v2-lora.md'>card</a></td></tr><tr><td>qwen3-1.7b-ben-franklin-thinking-v3-negative-identity-lora</td><td>Qwen3 1.7B 4-bit</td><td>0.78</td><td>32</td><td></td><td></td><td><a href='model_cards/qwen3-1.7b-ben-franklin-thinking-v3-negative-identity-lora.md'>card</a></td></tr><tr><td>qwen3-1.7b-ben-franklin-thinking-v4-balanced-lora</td><td>Qwen3 1.7B 4-bit</td><td>0.78</td><td>32</td><td></td><td></td><td><a href='model_cards/qwen3-1.7b-ben-franklin-thinking-v4-balanced-lora.md'>card</a></td></tr><tr><td>qwen3-1.7b-ben-franklin-thinking-v5-ood-fixed-lora</td><td>Qwen3 1.7B 4-bit</td><td>0.78</td><td>32</td><td></td><td></td><td><a href='model_cards/qwen3-1.7b-ben-franklin-thinking-v5-ood-fixed-lora.md'>card</a></td></tr><tr><td>qwen3-1.7b-ben-franklin-thinking-v6-1-ood-fixed-lora</td><td>Qwen3 1.7B 4-bit</td><td>0.99</td><td>32</td><td></td><td></td><td><a href='model_cards/qwen3-1.7b-ben-franklin-thinking-v6-1-ood-fixed-lora.md'>card</a></td></tr><tr><td>qwen3-1.7b-ben-franklin-thinking-v6-contrastive-lora</td><td>Qwen3 1.7B 4-bit</td><td>1.62</td><td>32</td><td></td><td></td><td><a href='model_cards/qwen3-1.7b-ben-franklin-thinking-v6-contrastive-lora.md'>card</a></td></tr><tr><td>qwen3-1.7b-ben-franklin-thinking-v7-natural-dialogue-lora</td><td>Qwen3 1.7B 4-bit</td><td>1.2</td><td>32</td><td></td><td></td><td><a href='model_cards/qwen3-1.7b-ben-franklin-thinking-v7-natural-dialogue-lora.md'>card</a></td></tr><tr><td>qwen3-1.7b-ben-franklin-thinking-v8-minimal-thought-lora</td><td>Qwen3 1.7B 4-bit</td><td>0.99</td><td>32</td><td></td><td></td><td><a href='model_cards/qwen3-1.7b-ben-franklin-thinking-v8-minimal-thought-lora.md'>card</a></td></tr><tr><td>qwen3-1.7b-ben-franklin-thinking-v9-factual-dialogue-lora</td><td>Qwen3 1.7B 4-bit</td><td>1.62</td><td>32</td><td></td><td></td><td><a href='model_cards/qwen3-1.7b-ben-franklin-thinking-v9-factual-dialogue-lora.md'>card</a></td></tr><tr><td>qwen3-1.7b-ben-franklin-thinking-v9-from-v2-factual-dialogue-lora</td><td>Qwen3 1.7B 4-bit</td><td>1.62</td><td>32</td><td></td><td></td><td><a href='model_cards/qwen3-1.7b-ben-franklin-thinking-v9-from-v2-factual-dialogue-lora.md'>card</a></td></tr></table></section> |
| <section class='panel'><h2>Data mix overview</h2><p>Training data included persona SFT, OpenAI-expanded Franklin dialogue, thinking/identity reinforcement, OOD corrections, natural dialogue repair, tool-call cleanup, English-lock cleanup, 7B ChatML answer-clean rows, and targeted coherence/factual repair rows.</p><table><tr><th>Dataset</th><th>Rows</th></tr><tr><td>franklin_7b_coherence_repair_v2.jsonl</td><td>287</td></tr><tr><td>franklin_7b_factual_coherence_repair_v3.jsonl</td><td>308</td></tr><tr><td>franklin_identity_reinforcement.jsonl</td><td>212</td></tr><tr><td>franklin_negative_identity_thinking.jsonl</td><td>829</td></tr><tr><td>franklin_persona_openai_expanded.jsonl</td><td>891</td></tr><tr><td>franklin_persona_sft.jsonl</td><td>288</td></tr><tr><td>franklin_qwen3_4b_answer_only.jsonl</td><td>3253</td></tr><tr><td>franklin_qwen3_4b_english_lock_cleanup.jsonl</td><td>816</td></tr><tr><td>franklin_qwen3_4b_toolcall_cleanup.jsonl</td><td>920</td></tr><tr><td>franklin_qwen3_8b_chatml_answer_clean.jsonl</td><td>2400</td></tr><tr><td>franklin_thinking_sft.jsonl</td><td>1259</td></tr><tr><td>franklin_thinking_strong_reinforcement.jsonl</td><td>575</td></tr><tr><td>franklin_v4_general_balanced_thinking.jsonl</td><td>825</td></tr><tr><td>franklin_v5_out_of_domain_correction.jsonl</td><td>684</td></tr><tr><td>franklin_v6_1_targeted_ood_fix.jsonl</td><td>870</td></tr><tr><td>franklin_v6_contrastive_thinking.jsonl</td><td>599</td></tr><tr><td>franklin_v6_contrastive_thinking_final.jsonl</td><td>2684</td></tr><tr><td>franklin_v6_contrastive_thinking_weighted.jsonl</td><td>1584</td></tr><tr><td>franklin_v7_natural_dialogue_repair.jsonl</td><td>1830</td></tr><tr><td>franklin_v8_minimal_thought_repair.jsonl</td><td>1040</td></tr><tr><td>franklin_v9_factual_dialogue.jsonl</td><td>3019</td></tr></table></section> |
| <section class='panel'><h2>How these models might be useful</h2><ul><li>Local historical-character chatbots with a warmer Franklin voice.</li><li>RAG-backed educational demos where retrieved facts ground the persona.</li><li>Comparative LoRA experiments on identity persistence, tool-call cleanup, English-only steering, and natural dialogue.</li><li>Small-footprint offline agents that offer practical advice in a civic/philosophical style.</li><li>Training-data research: the folder preserves adapter artifacts, configs, model cards, and benchmark links for future ablations.</li></ul></section> |
|
|
|
|
|
|
| <section class='panel'><h2>Browse models</h2><input id='q' placeholder='Filter by name, base family, strength, weakness, data mix...' oninput='filterCards()'></section> |
|
|
| <section class='model-card' data-family='Qwen2.5 7B Instruct 4-bit' data-name='qwen2.5-7b-ben-franklin-v1-lite-r4-qv'> |
| <h3>qwen2.5-7b-ben-franklin-v1-lite-r4-qv</h3> |
| <p><b>Family:</b> Qwen2.5 7B Instruct 4-bit · <b>Size:</b> 0.06 GB · <b>LoRA:</b> r=4 alpha=8 · <a href='model_cards/qwen2.5-7b-ben-franklin-v1-lite-r4-qv.md'>model card</a></p> |
| <p><b>Benchmark:</b> 30 · <b>Flags:</b> {'base_identity_leak': 1, 'overdenial': 1}</p> |
| <div class='cols'><div><h4>Strengths</h4><ul><li>Largest Benjamin Franklin LoRA family proven trainable on this RTX 3070 8GB machine.</li><li>Best base reasoning/coherence among the local Franklin adapters.</li><li>Good short-turn continuity in the coherence benchmark.</li><li>Benchmark score: 30 with flags {'base_identity_leak': 1, 'overdenial': 1}.</li></ul></div><div><h4>Weaknesses</h4><ul><li>Qwen2.5 7B base is already highly steerable, so improvements over the prompted base are modest.</li><li>q/v-only LoRA is weak for implanting stubborn factual corrections.</li><li>Craven Street/Hewson factuality remains unreliable unless retrieval/prompt context is supplied.</li></ul></div></div> |
| <p><b>Data mix:</b><br>franklin_qwen3_8b_chatml_answer_clean.jsonl: 2400</p> |
| <p><b>Compute:</b> 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.</p> |
| <code>./adapters/qwen2.5-7b-ben-franklin-v1-lite-r4-qv</code> |
| </section> |
| <section class='model-card' data-family='Qwen2.5 7B Instruct 4-bit' data-name='qwen2.5-7b-ben-franklin-v2-coherence-r4-qv'> |
| <h3>qwen2.5-7b-ben-franklin-v2-coherence-r4-qv</h3> |
| <p><b>Family:</b> Qwen2.5 7B Instruct 4-bit · <b>Size:</b> 0.06 GB · <b>LoRA:</b> r=4 alpha=8 · <a href='model_cards/qwen2.5-7b-ben-franklin-v2-coherence-r4-qv.md'>model card</a></p> |
| <p><b>Benchmark:</b> 33 · <b>Flags:</b> {'base_identity_leak': 1}</p> |
| <div class='cols'><div><h4>Strengths</h4><ul><li>Largest Benjamin Franklin LoRA family proven trainable on this RTX 3070 8GB machine.</li><li>Best base reasoning/coherence among the local Franklin adapters.</li><li>Good short-turn continuity in the coherence benchmark.</li><li>Best modest improvement over 7B v1 for coherence and reduced over-denial in the broad benchmark.</li></ul></div><div><h4>Weaknesses</h4><ul><li>Qwen2.5 7B base is already highly steerable, so improvements over the prompted base are modest.</li><li>q/v-only LoRA is weak for implanting stubborn factual corrections.</li><li>Craven Street/Hewson factuality remains unreliable unless retrieval/prompt context is supplied.</li></ul></div></div> |
| <p><b>Data mix:</b><br>franklin_qwen3_8b_chatml_answer_clean.jsonl: 2400<br>franklin_7b_coherence_repair_v2.jsonl: 287</p> |
| <p><b>Compute:</b> 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.</p> |
| <code>./adapters/qwen2.5-7b-ben-franklin-v2-coherence-r4-qv</code> |
| </section> |
| <section class='model-card' data-family='Qwen2.5 7B Instruct 4-bit' data-name='qwen2.5-7b-ben-franklin-v3-factual-coherence-r4-qv'> |
| <h3>qwen2.5-7b-ben-franklin-v3-factual-coherence-r4-qv</h3> |
| <p><b>Family:</b> Qwen2.5 7B Instruct 4-bit · <b>Size:</b> 0.06 GB · <b>LoRA:</b> r=4 alpha=8 · <a href='model_cards/qwen2.5-7b-ben-franklin-v3-factual-coherence-r4-qv.md'>model card</a></p> |
| <p><b>Benchmark:</b> 35 · <b>Flags:</b> {'base_identity_leak': 1}</p> |
| <div class='cols'><div><h4>Strengths</h4><ul><li>Largest Benjamin Franklin LoRA family proven trainable on this RTX 3070 8GB machine.</li><li>Best base reasoning/coherence among the local Franklin adapters.</li><li>Good short-turn continuity in the coherence benchmark.</li><li>Best numeric coherence benchmark score among evaluated 7B adapters.</li></ul></div><div><h4>Weaknesses</h4><ul><li>Qwen2.5 7B base is already highly steerable, so improvements over the prompted base are modest.</li><li>q/v-only LoRA is weak for implanting stubborn factual corrections.</li><li>Craven Street/Hewson factuality remains unreliable unless retrieval/prompt context is supplied.</li><li>Despite the name, not a clean factual fix: Craven Street answer still hallucinated.</li></ul></div></div> |
| <p><b>Data mix:</b><br>franklin_qwen3_8b_chatml_answer_clean.jsonl: 2400<br>franklin_7b_coherence_repair_v2.jsonl: 287<br>franklin_7b_factual_coherence_repair_v3.jsonl: 308</p> |
| <p><b>Compute:</b> 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.</p> |
| <code>./adapters/qwen2.5-7b-ben-franklin-v3-factual-coherence-r4-qv</code> |
| </section> |
| <section class='model-card' data-family='Qwen2.5 7B Instruct 4-bit' data-name='qwen2.5-7b-ben-franklin-v3c-factual-r8-qv-from-base'> |
| <h3>qwen2.5-7b-ben-franklin-v3c-factual-r8-qv-from-base</h3> |
| <p><b>Family:</b> Qwen2.5 7B Instruct 4-bit · <b>Size:</b> 0.08 GB · <b>LoRA:</b> r=8 alpha=16 · <a href='model_cards/qwen2.5-7b-ben-franklin-v3c-factual-r8-qv-from-base.md'>model card</a></p> |
| <p><b>Benchmark:</b> 28 · <b>Flags:</b> {'base_identity_leak': 1, 'overdenial': 1}</p> |
| <div class='cols'><div><h4>Strengths</h4><ul><li>Largest Benjamin Franklin LoRA family proven trainable on this RTX 3070 8GB machine.</li><li>Best base reasoning/coherence among the local Franklin adapters.</li><li>Good short-turn continuity in the coherence benchmark.</li><li>Proves r=8 q/v LoRA from clean 7B base can train on this 8GB GPU.</li></ul></div><div><h4>Weaknesses</h4><ul><li>Qwen2.5 7B base is already highly steerable, so improvements over the prompted base are modest.</li><li>q/v-only LoRA is weak for implanting stubborn factual corrections.</li><li>Craven Street/Hewson factuality remains unreliable unless retrieval/prompt context is supplied.</li><li>Regressed versus v1/v2/v3 in the coherence benchmark.</li></ul></div></div> |
| <p><b>Data mix:</b><br>franklin_qwen3_8b_chatml_answer_clean.jsonl: 2400<br>franklin_7b_factual_coherence_repair_v3.jsonl: 308</p> |
| <p><b>Compute:</b> 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.</p> |
| <code>./adapters/qwen2.5-7b-ben-franklin-v3c-factual-r8-qv-from-base</code> |
| </section> |
| <section class='model-card' data-family='Qwen3 4B Instruct 4-bit' data-name='qwen3-4b-instruct-2507-ben-franklin-v1-lora'> |
| <h3>qwen3-4b-instruct-2507-ben-franklin-v1-lora</h3> |
| <p><b>Family:</b> Qwen3 4B Instruct 4-bit · <b>Size:</b> 1.34 GB · <b>LoRA:</b> r=16 alpha=32 · <a href='model_cards/qwen3-4b-instruct-2507-ben-franklin-v1-lora.md'>model card</a></p> |
| <p><b>Benchmark:</b> · <b>Flags:</b> not benchmarked</p> |
| <div class='cols'><div><h4>Strengths</h4><ul><li>Middle-size family: more capable than 1.7B while still comfortable on 8GB VRAM.</li><li>Several variants target ChatML/completion formatting, tool-call cleanup, and English-lock behavior.</li></ul></div><div><h4>Weaknesses</h4><ul><li>Some later 4B adapters, especially v5, know targeted facts but emit visible tool_call tags offline.</li><li>Can leak base-model identity or policy/meta phrasing depending on prompt path.</li></ul></div></div> |
| <p><b>Data mix:</b><br>franklin_qwen3_4b_answer_only.jsonl: 3253</p> |
| <p><b>Compute:</b> Inference/training: comfortable on RTX 3070 8GB in 4-bit. Full-module LoRA at r=16-32 was used historically; expect several GB VRAM and slower but practical training.</p> |
| <code>./adapters/qwen3-4b-instruct-2507-ben-franklin-v1-lora</code> |
| </section> |
| <section class='model-card' data-family='Qwen3 4B Instruct 4-bit' data-name='qwen3-4b-instruct-2507-ben-franklin-v2-chatml-lora'> |
| <h3>qwen3-4b-instruct-2507-ben-franklin-v2-chatml-lora</h3> |
| <p><b>Family:</b> Qwen3 4B Instruct 4-bit · <b>Size:</b> 2.19 GB · <b>LoRA:</b> r=32 alpha=64 · <a href='model_cards/qwen3-4b-instruct-2507-ben-franklin-v2-chatml-lora.md'>model card</a></p> |
| <p><b>Benchmark:</b> · <b>Flags:</b> not benchmarked</p> |
| <div class='cols'><div><h4>Strengths</h4><ul><li>Middle-size family: more capable than 1.7B while still comfortable on 8GB VRAM.</li><li>Several variants target ChatML/completion formatting, tool-call cleanup, and English-lock behavior.</li></ul></div><div><h4>Weaknesses</h4><ul><li>Some later 4B adapters, especially v5, know targeted facts but emit visible tool_call tags offline.</li><li>Can leak base-model identity or policy/meta phrasing depending on prompt path.</li></ul></div></div> |
| <p><b>Data mix:</b><br>franklin_qwen3_4b_answer_only.jsonl: 3253</p> |
| <p><b>Compute:</b> Inference/training: comfortable on RTX 3070 8GB in 4-bit. Full-module LoRA at r=16-32 was used historically; expect several GB VRAM and slower but practical training.</p> |
| <code>./adapters/qwen3-4b-instruct-2507-ben-franklin-v2-chatml-lora</code> |
| </section> |
| <section class='model-card' data-family='Qwen3 4B Instruct 4-bit' data-name='qwen3-4b-instruct-2507-ben-franklin-v3-chatml-completions-lora'> |
| <h3>qwen3-4b-instruct-2507-ben-franklin-v3-chatml-completions-lora</h3> |
| <p><b>Family:</b> Qwen3 4B Instruct 4-bit · <b>Size:</b> 2.19 GB · <b>LoRA:</b> r=32 alpha=64 · <a href='model_cards/qwen3-4b-instruct-2507-ben-franklin-v3-chatml-completions-lora.md'>model card</a></p> |
| <p><b>Benchmark:</b> · <b>Flags:</b> not benchmarked</p> |
| <div class='cols'><div><h4>Strengths</h4><ul><li>Middle-size family: more capable than 1.7B while still comfortable on 8GB VRAM.</li><li>Several variants target ChatML/completion formatting, tool-call cleanup, and English-lock behavior.</li></ul></div><div><h4>Weaknesses</h4><ul><li>Some later 4B adapters, especially v5, know targeted facts but emit visible tool_call tags offline.</li><li>Can leak base-model identity or policy/meta phrasing depending on prompt path.</li></ul></div></div> |
| <p><b>Data mix:</b><br>franklin_qwen3_4b_answer_only.jsonl: 3253</p> |
| <p><b>Compute:</b> Inference/training: comfortable on RTX 3070 8GB in 4-bit. Full-module LoRA at r=16-32 was used historically; expect several GB VRAM and slower but practical training.</p> |
| <code>./adapters/qwen3-4b-instruct-2507-ben-franklin-v3-chatml-completions-lora</code> |
| </section> |
| <section class='model-card' data-family='Qwen3 4B Instruct 4-bit' data-name='qwen3-4b-instruct-2507-ben-franklin-v4-toolcall-clean-lora'> |
| <h3>qwen3-4b-instruct-2507-ben-franklin-v4-toolcall-clean-lora</h3> |
| <p><b>Family:</b> Qwen3 4B Instruct 4-bit · <b>Size:</b> 1.81 GB · <b>LoRA:</b> r=32 alpha=64 · <a href='model_cards/qwen3-4b-instruct-2507-ben-franklin-v4-toolcall-clean-lora.md'>model card</a></p> |
| <p><b>Benchmark:</b> · <b>Flags:</b> not benchmarked</p> |
| <div class='cols'><div><h4>Strengths</h4><ul><li>Middle-size family: more capable than 1.7B while still comfortable on 8GB VRAM.</li><li>Several variants target ChatML/completion formatting, tool-call cleanup, and English-lock behavior.</li><li>Targeted cleanup of visible tool_call artifacts.</li></ul></div><div><h4>Weaknesses</h4><ul><li>Some later 4B adapters, especially v5, know targeted facts but emit visible tool_call tags offline.</li><li>Can leak base-model identity or policy/meta phrasing depending on prompt path.</li></ul></div></div> |
| <p><b>Data mix:</b><br>franklin_qwen3_4b_toolcall_cleanup.jsonl: 920</p> |
| <p><b>Compute:</b> Inference/training: comfortable on RTX 3070 8GB in 4-bit. Full-module LoRA at r=16-32 was used historically; expect several GB VRAM and slower but practical training.</p> |
| <code>./adapters/qwen3-4b-instruct-2507-ben-franklin-v4-toolcall-clean-lora</code> |
| </section> |
| <section class='model-card' data-family='Qwen3 4B Instruct 4-bit' data-name='qwen3-4b-instruct-2507-ben-franklin-v5-english-lock-lora'> |
| <h3>qwen3-4b-instruct-2507-ben-franklin-v5-english-lock-lora</h3> |
| <p><b>Family:</b> Qwen3 4B Instruct 4-bit · <b>Size:</b> 1.81 GB · <b>LoRA:</b> r=32 alpha=64 · <a href='model_cards/qwen3-4b-instruct-2507-ben-franklin-v5-english-lock-lora.md'>model card</a></p> |
| <p><b>Benchmark:</b> -92 · <b>Flags:</b> {'tool_call': 15, 'continuity_miss': 1}</p> |
| <div class='cols'><div><h4>Strengths</h4><ul><li>Middle-size family: more capable than 1.7B while still comfortable on 8GB VRAM.</li><li>Several variants target ChatML/completion formatting, tool-call cleanup, and English-lock behavior.</li><li>Contains useful cleaned English/factual phrasing, including better Craven/Hewson material.</li><li>Benchmark score: -92 with flags {'tool_call': 15, 'continuity_miss': 1}.</li></ul></div><div><h4>Weaknesses</h4><ul><li>Some later 4B adapters, especially v5, know targeted facts but emit visible tool_call tags offline.</li><li>Can leak base-model identity or policy/meta phrasing depending on prompt path.</li><li>Offline benchmark showed severe visible tool_call tag regression.</li></ul></div></div> |
| <p><b>Data mix:</b><br>franklin_qwen3_4b_english_lock_cleanup.jsonl: 816</p> |
| <p><b>Compute:</b> Inference/training: comfortable on RTX 3070 8GB in 4-bit. Full-module LoRA at r=16-32 was used historically; expect several GB VRAM and slower but practical training.</p> |
| <code>./adapters/qwen3-4b-instruct-2507-ben-franklin-v5-english-lock-lora</code> |
| </section> |
| <section class='model-card' data-family='Qwen3 1.7B 4-bit' data-name='qwen3-1.7b-ben-franklin-identity-reinforced-lora'> |
| <h3>qwen3-1.7b-ben-franklin-identity-reinforced-lora</h3> |
| <p><b>Family:</b> Qwen3 1.7B 4-bit · <b>Size:</b> 0.99 GB · <b>LoRA:</b> r=32 alpha=64 · <a href='model_cards/qwen3-1.7b-ben-franklin-identity-reinforced-lora.md'>model card</a></p> |
| <p><b>Benchmark:</b> · <b>Flags:</b> not benchmarked</p> |
| <div class='cols'><div><h4>Strengths</h4><ul><li>Largest Benjamin Franklin LoRA family proven trainable on this RTX 3070 8GB machine.</li><li>Best base reasoning/coherence among the local Franklin adapters.</li><li>Good short-turn continuity in the coherence benchmark.</li></ul></div><div><h4>Weaknesses</h4><ul><li>Qwen2.5 7B base is already highly steerable, so improvements over the prompted base are modest.</li><li>q/v-only LoRA is weak for implanting stubborn factual corrections.</li><li>Craven Street/Hewson factuality remains unreliable unless retrieval/prompt context is supplied.</li></ul></div></div> |
| <p><b>Data mix:</b><br>franklin_qwen3_8b_chatml_answer_clean.jsonl: 2400</p> |
| <p><b>Compute:</b> 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.</p> |
| <code>./adapters/qwen3-1.7b-ben-franklin-identity-reinforced-lora</code> |
| </section> |
| <section class='model-card' data-family='Qwen3 1.7B 4-bit' data-name='qwen3-1.7b-ben-franklin-openai-expanded-lora'> |
| <h3>qwen3-1.7b-ben-franklin-openai-expanded-lora</h3> |
| <p><b>Family:</b> Qwen3 1.7B 4-bit · <b>Size:</b> 1.2 GB · <b>LoRA:</b> r=32 alpha=64 · <a href='model_cards/qwen3-1.7b-ben-franklin-openai-expanded-lora.md'>model card</a></p> |
| <p><b>Benchmark:</b> · <b>Flags:</b> not benchmarked</p> |
| <div class='cols'><div><h4>Strengths</h4><ul><li>Largest Benjamin Franklin LoRA family proven trainable on this RTX 3070 8GB machine.</li><li>Best base reasoning/coherence among the local Franklin adapters.</li><li>Good short-turn continuity in the coherence benchmark.</li></ul></div><div><h4>Weaknesses</h4><ul><li>Qwen2.5 7B base is already highly steerable, so improvements over the prompted base are modest.</li><li>q/v-only LoRA is weak for implanting stubborn factual corrections.</li><li>Craven Street/Hewson factuality remains unreliable unless retrieval/prompt context is supplied.</li></ul></div></div> |
| <p><b>Data mix:</b><br>franklin_qwen3_8b_chatml_answer_clean.jsonl: 2400</p> |
| <p><b>Compute:</b> 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.</p> |
| <code>./adapters/qwen3-1.7b-ben-franklin-openai-expanded-lora</code> |
| </section> |
| <section class='model-card' data-family='Qwen3 1.7B 4-bit' data-name='qwen3-1.7b-ben-franklin-thinking-lora'> |
| <h3>qwen3-1.7b-ben-franklin-thinking-lora</h3> |
| <p><b>Family:</b> Qwen3 1.7B 4-bit · <b>Size:</b> 0.99 GB · <b>LoRA:</b> r=32 alpha=64 · <a href='model_cards/qwen3-1.7b-ben-franklin-thinking-lora.md'>model card</a></p> |
| <p><b>Benchmark:</b> · <b>Flags:</b> not benchmarked</p> |
| <div class='cols'><div><h4>Strengths</h4><ul><li>Largest Benjamin Franklin LoRA family proven trainable on this RTX 3070 8GB machine.</li><li>Best base reasoning/coherence among the local Franklin adapters.</li><li>Good short-turn continuity in the coherence benchmark.</li></ul></div><div><h4>Weaknesses</h4><ul><li>Qwen2.5 7B base is already highly steerable, so improvements over the prompted base are modest.</li><li>q/v-only LoRA is weak for implanting stubborn factual corrections.</li><li>Craven Street/Hewson factuality remains unreliable unless retrieval/prompt context is supplied.</li></ul></div></div> |
| <p><b>Data mix:</b><br>franklin_qwen3_8b_chatml_answer_clean.jsonl: 2400</p> |
| <p><b>Compute:</b> 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.</p> |
| <code>./adapters/qwen3-1.7b-ben-franklin-thinking-lora</code> |
| </section> |
| <section class='model-card' data-family='Qwen3 1.7B 4-bit' data-name='qwen3-1.7b-ben-franklin-thinking-v2-lora'> |
| <h3>qwen3-1.7b-ben-franklin-thinking-v2-lora</h3> |
| <p><b>Family:</b> Qwen3 1.7B 4-bit · <b>Size:</b> 0.78 GB · <b>LoRA:</b> r=32 alpha=64 · <a href='model_cards/qwen3-1.7b-ben-franklin-thinking-v2-lora.md'>model card</a></p> |
| <p><b>Benchmark:</b> · <b>Flags:</b> not benchmarked</p> |
| <div class='cols'><div><h4>Strengths</h4><ul><li>Largest Benjamin Franklin LoRA family proven trainable on this RTX 3070 8GB machine.</li><li>Best base reasoning/coherence among the local Franklin adapters.</li><li>Good short-turn continuity in the coherence benchmark.</li></ul></div><div><h4>Weaknesses</h4><ul><li>Qwen2.5 7B base is already highly steerable, so improvements over the prompted base are modest.</li><li>q/v-only LoRA is weak for implanting stubborn factual corrections.</li><li>Craven Street/Hewson factuality remains unreliable unless retrieval/prompt context is supplied.</li></ul></div></div> |
| <p><b>Data mix:</b><br>franklin_qwen3_8b_chatml_answer_clean.jsonl: 2400<br>franklin_7b_coherence_repair_v2.jsonl: 287</p> |
| <p><b>Compute:</b> 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.</p> |
| <code>./adapters/qwen3-1.7b-ben-franklin-thinking-v2-lora</code> |
| </section> |
| <section class='model-card' data-family='Qwen3 1.7B 4-bit' data-name='qwen3-1.7b-ben-franklin-thinking-v3-negative-identity-lora'> |
| <h3>qwen3-1.7b-ben-franklin-thinking-v3-negative-identity-lora</h3> |
| <p><b>Family:</b> Qwen3 1.7B 4-bit · <b>Size:</b> 0.78 GB · <b>LoRA:</b> r=32 alpha=64 · <a href='model_cards/qwen3-1.7b-ben-franklin-thinking-v3-negative-identity-lora.md'>model card</a></p> |
| <p><b>Benchmark:</b> · <b>Flags:</b> not benchmarked</p> |
| <div class='cols'><div><h4>Strengths</h4><ul><li>Largest Benjamin Franklin LoRA family proven trainable on this RTX 3070 8GB machine.</li><li>Best base reasoning/coherence among the local Franklin adapters.</li><li>Good short-turn continuity in the coherence benchmark.</li></ul></div><div><h4>Weaknesses</h4><ul><li>Qwen2.5 7B base is already highly steerable, so improvements over the prompted base are modest.</li><li>q/v-only LoRA is weak for implanting stubborn factual corrections.</li><li>Craven Street/Hewson factuality remains unreliable unless retrieval/prompt context is supplied.</li></ul></div></div> |
| <p><b>Data mix:</b><br>franklin_qwen3_8b_chatml_answer_clean.jsonl: 2400<br>franklin_7b_factual_coherence_repair_v3.jsonl: 308</p> |
| <p><b>Compute:</b> 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.</p> |
| <code>./adapters/qwen3-1.7b-ben-franklin-thinking-v3-negative-identity-lora</code> |
| </section> |
| <section class='model-card' data-family='Qwen3 1.7B 4-bit' data-name='qwen3-1.7b-ben-franklin-thinking-v4-balanced-lora'> |
| <h3>qwen3-1.7b-ben-franklin-thinking-v4-balanced-lora</h3> |
| <p><b>Family:</b> Qwen3 1.7B 4-bit · <b>Size:</b> 0.78 GB · <b>LoRA:</b> r=32 alpha=64 · <a href='model_cards/qwen3-1.7b-ben-franklin-thinking-v4-balanced-lora.md'>model card</a></p> |
| <p><b>Benchmark:</b> · <b>Flags:</b> not benchmarked</p> |
| <div class='cols'><div><h4>Strengths</h4><ul><li>Largest Benjamin Franklin LoRA family proven trainable on this RTX 3070 8GB machine.</li><li>Best base reasoning/coherence among the local Franklin adapters.</li><li>Good short-turn continuity in the coherence benchmark.</li></ul></div><div><h4>Weaknesses</h4><ul><li>Qwen2.5 7B base is already highly steerable, so improvements over the prompted base are modest.</li><li>q/v-only LoRA is weak for implanting stubborn factual corrections.</li><li>Craven Street/Hewson factuality remains unreliable unless retrieval/prompt context is supplied.</li></ul></div></div> |
| <p><b>Data mix:</b><br>franklin_qwen3_8b_chatml_answer_clean.jsonl: 2400</p> |
| <p><b>Compute:</b> 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.</p> |
| <code>./adapters/qwen3-1.7b-ben-franklin-thinking-v4-balanced-lora</code> |
| </section> |
| <section class='model-card' data-family='Qwen3 1.7B 4-bit' data-name='qwen3-1.7b-ben-franklin-thinking-v5-ood-fixed-lora'> |
| <h3>qwen3-1.7b-ben-franklin-thinking-v5-ood-fixed-lora</h3> |
| <p><b>Family:</b> Qwen3 1.7B 4-bit · <b>Size:</b> 0.78 GB · <b>LoRA:</b> r=32 alpha=64 · <a href='model_cards/qwen3-1.7b-ben-franklin-thinking-v5-ood-fixed-lora.md'>model card</a></p> |
| <p><b>Benchmark:</b> · <b>Flags:</b> not benchmarked</p> |
| <div class='cols'><div><h4>Strengths</h4><ul><li>Largest Benjamin Franklin LoRA family proven trainable on this RTX 3070 8GB machine.</li><li>Best base reasoning/coherence among the local Franklin adapters.</li><li>Good short-turn continuity in the coherence benchmark.</li></ul></div><div><h4>Weaknesses</h4><ul><li>Qwen2.5 7B base is already highly steerable, so improvements over the prompted base are modest.</li><li>q/v-only LoRA is weak for implanting stubborn factual corrections.</li><li>Craven Street/Hewson factuality remains unreliable unless retrieval/prompt context is supplied.</li></ul></div></div> |
| <p><b>Data mix:</b><br>franklin_qwen3_8b_chatml_answer_clean.jsonl: 2400</p> |
| <p><b>Compute:</b> 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.</p> |
| <code>./adapters/qwen3-1.7b-ben-franklin-thinking-v5-ood-fixed-lora</code> |
| </section> |
| <section class='model-card' data-family='Qwen3 1.7B 4-bit' data-name='qwen3-1.7b-ben-franklin-thinking-v6-1-ood-fixed-lora'> |
| <h3>qwen3-1.7b-ben-franklin-thinking-v6-1-ood-fixed-lora</h3> |
| <p><b>Family:</b> Qwen3 1.7B 4-bit · <b>Size:</b> 0.99 GB · <b>LoRA:</b> r=32 alpha=64 · <a href='model_cards/qwen3-1.7b-ben-franklin-thinking-v6-1-ood-fixed-lora.md'>model card</a></p> |
| <p><b>Benchmark:</b> · <b>Flags:</b> not benchmarked</p> |
| <div class='cols'><div><h4>Strengths</h4><ul><li>Largest Benjamin Franklin LoRA family proven trainable on this RTX 3070 8GB machine.</li><li>Best base reasoning/coherence among the local Franklin adapters.</li><li>Good short-turn continuity in the coherence benchmark.</li></ul></div><div><h4>Weaknesses</h4><ul><li>Qwen2.5 7B base is already highly steerable, so improvements over the prompted base are modest.</li><li>q/v-only LoRA is weak for implanting stubborn factual corrections.</li><li>Craven Street/Hewson factuality remains unreliable unless retrieval/prompt context is supplied.</li></ul></div></div> |
| <p><b>Data mix:</b><br>franklin_qwen3_8b_chatml_answer_clean.jsonl: 2400</p> |
| <p><b>Compute:</b> 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.</p> |
| <code>./adapters/qwen3-1.7b-ben-franklin-thinking-v6-1-ood-fixed-lora</code> |
| </section> |
| <section class='model-card' data-family='Qwen3 1.7B 4-bit' data-name='qwen3-1.7b-ben-franklin-thinking-v6-contrastive-lora'> |
| <h3>qwen3-1.7b-ben-franklin-thinking-v6-contrastive-lora</h3> |
| <p><b>Family:</b> Qwen3 1.7B 4-bit · <b>Size:</b> 1.62 GB · <b>LoRA:</b> r=32 alpha=64 · <a href='model_cards/qwen3-1.7b-ben-franklin-thinking-v6-contrastive-lora.md'>model card</a></p> |
| <p><b>Benchmark:</b> · <b>Flags:</b> not benchmarked</p> |
| <div class='cols'><div><h4>Strengths</h4><ul><li>Largest Benjamin Franklin LoRA family proven trainable on this RTX 3070 8GB machine.</li><li>Best base reasoning/coherence among the local Franklin adapters.</li><li>Good short-turn continuity in the coherence benchmark.</li></ul></div><div><h4>Weaknesses</h4><ul><li>Qwen2.5 7B base is already highly steerable, so improvements over the prompted base are modest.</li><li>q/v-only LoRA is weak for implanting stubborn factual corrections.</li><li>Craven Street/Hewson factuality remains unreliable unless retrieval/prompt context is supplied.</li></ul></div></div> |
| <p><b>Data mix:</b><br>franklin_qwen3_8b_chatml_answer_clean.jsonl: 2400</p> |
| <p><b>Compute:</b> 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.</p> |
| <code>./adapters/qwen3-1.7b-ben-franklin-thinking-v6-contrastive-lora</code> |
| </section> |
| <section class='model-card' data-family='Qwen3 1.7B 4-bit' data-name='qwen3-1.7b-ben-franklin-thinking-v7-natural-dialogue-lora'> |
| <h3>qwen3-1.7b-ben-franklin-thinking-v7-natural-dialogue-lora</h3> |
| <p><b>Family:</b> Qwen3 1.7B 4-bit · <b>Size:</b> 1.2 GB · <b>LoRA:</b> r=32 alpha=64 · <a href='model_cards/qwen3-1.7b-ben-franklin-thinking-v7-natural-dialogue-lora.md'>model card</a></p> |
| <p><b>Benchmark:</b> · <b>Flags:</b> not benchmarked</p> |
| <div class='cols'><div><h4>Strengths</h4><ul><li>Largest Benjamin Franklin LoRA family proven trainable on this RTX 3070 8GB machine.</li><li>Best base reasoning/coherence among the local Franklin adapters.</li><li>Good short-turn continuity in the coherence benchmark.</li><li>Focused on more natural short conversational replies.</li></ul></div><div><h4>Weaknesses</h4><ul><li>Qwen2.5 7B base is already highly steerable, so improvements over the prompted base are modest.</li><li>q/v-only LoRA is weak for implanting stubborn factual corrections.</li><li>Craven Street/Hewson factuality remains unreliable unless retrieval/prompt context is supplied.</li></ul></div></div> |
| <p><b>Data mix:</b><br>franklin_qwen3_8b_chatml_answer_clean.jsonl: 2400</p> |
| <p><b>Compute:</b> 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.</p> |
| <code>./adapters/qwen3-1.7b-ben-franklin-thinking-v7-natural-dialogue-lora</code> |
| </section> |
| <section class='model-card' data-family='Qwen3 1.7B 4-bit' data-name='qwen3-1.7b-ben-franklin-thinking-v8-minimal-thought-lora'> |
| <h3>qwen3-1.7b-ben-franklin-thinking-v8-minimal-thought-lora</h3> |
| <p><b>Family:</b> Qwen3 1.7B 4-bit · <b>Size:</b> 0.99 GB · <b>LoRA:</b> r=32 alpha=64 · <a href='model_cards/qwen3-1.7b-ben-franklin-thinking-v8-minimal-thought-lora.md'>model card</a></p> |
| <p><b>Benchmark:</b> · <b>Flags:</b> not benchmarked</p> |
| <div class='cols'><div><h4>Strengths</h4><ul><li>Largest Benjamin Franklin LoRA family proven trainable on this RTX 3070 8GB machine.</li><li>Best base reasoning/coherence among the local Franklin adapters.</li><li>Good short-turn continuity in the coherence benchmark.</li><li>Focused on reducing visible thought/over-reasoning style.</li></ul></div><div><h4>Weaknesses</h4><ul><li>Qwen2.5 7B base is already highly steerable, so improvements over the prompted base are modest.</li><li>q/v-only LoRA is weak for implanting stubborn factual corrections.</li><li>Craven Street/Hewson factuality remains unreliable unless retrieval/prompt context is supplied.</li></ul></div></div> |
| <p><b>Data mix:</b><br>franklin_qwen3_8b_chatml_answer_clean.jsonl: 2400</p> |
| <p><b>Compute:</b> 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.</p> |
| <code>./adapters/qwen3-1.7b-ben-franklin-thinking-v8-minimal-thought-lora</code> |
| </section> |
| <section class='model-card' data-family='Qwen3 1.7B 4-bit' data-name='qwen3-1.7b-ben-franklin-thinking-v9-factual-dialogue-lora'> |
| <h3>qwen3-1.7b-ben-franklin-thinking-v9-factual-dialogue-lora</h3> |
| <p><b>Family:</b> Qwen3 1.7B 4-bit · <b>Size:</b> 1.62 GB · <b>LoRA:</b> r=32 alpha=64 · <a href='model_cards/qwen3-1.7b-ben-franklin-thinking-v9-factual-dialogue-lora.md'>model card</a></p> |
| <p><b>Benchmark:</b> · <b>Flags:</b> not benchmarked</p> |
| <div class='cols'><div><h4>Strengths</h4><ul><li>Largest Benjamin Franklin LoRA family proven trainable on this RTX 3070 8GB machine.</li><li>Best base reasoning/coherence among the local Franklin adapters.</li><li>Good short-turn continuity in the coherence benchmark.</li><li>Focused on factual dialogue and hard Franklin biography prompts.</li></ul></div><div><h4>Weaknesses</h4><ul><li>Qwen2.5 7B base is already highly steerable, so improvements over the prompted base are modest.</li><li>q/v-only LoRA is weak for implanting stubborn factual corrections.</li><li>Craven Street/Hewson factuality remains unreliable unless retrieval/prompt context is supplied.</li></ul></div></div> |
| <p><b>Data mix:</b><br>franklin_qwen3_8b_chatml_answer_clean.jsonl: 2400<br>franklin_7b_factual_coherence_repair_v3.jsonl: 308</p> |
| <p><b>Compute:</b> 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.</p> |
| <code>./adapters/qwen3-1.7b-ben-franklin-thinking-v9-factual-dialogue-lora</code> |
| </section> |
| <section class='model-card' data-family='Qwen3 1.7B 4-bit' data-name='qwen3-1.7b-ben-franklin-thinking-v9-from-v2-factual-dialogue-lora'> |
| <h3>qwen3-1.7b-ben-franklin-thinking-v9-from-v2-factual-dialogue-lora</h3> |
| <p><b>Family:</b> Qwen3 1.7B 4-bit · <b>Size:</b> 1.62 GB · <b>LoRA:</b> r=32 alpha=64 · <a href='model_cards/qwen3-1.7b-ben-franklin-thinking-v9-from-v2-factual-dialogue-lora.md'>model card</a></p> |
| <p><b>Benchmark:</b> · <b>Flags:</b> not benchmarked</p> |
| <div class='cols'><div><h4>Strengths</h4><ul><li>Largest Benjamin Franklin LoRA family proven trainable on this RTX 3070 8GB machine.</li><li>Best base reasoning/coherence among the local Franklin adapters.</li><li>Good short-turn continuity in the coherence benchmark.</li><li>Focused on factual dialogue and hard Franklin biography prompts.</li></ul></div><div><h4>Weaknesses</h4><ul><li>Qwen2.5 7B base is already highly steerable, so improvements over the prompted base are modest.</li><li>q/v-only LoRA is weak for implanting stubborn factual corrections.</li><li>Craven Street/Hewson factuality remains unreliable unless retrieval/prompt context is supplied.</li></ul></div></div> |
| <p><b>Data mix:</b><br>franklin_qwen3_8b_chatml_answer_clean.jsonl: 2400<br>franklin_7b_coherence_repair_v2.jsonl: 287<br>franklin_7b_factual_coherence_repair_v3.jsonl: 308</p> |
| <p><b>Compute:</b> 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.</p> |
| <code>./adapters/qwen3-1.7b-ben-franklin-thinking-v9-from-v2-factual-dialogue-lora</code> |
| </section> |
| </main><script> |
| function filterCards(){let q=document.getElementById('q').value.toLowerCase();for(let c of document.querySelectorAll('.model-card')){c.style.display=c.innerText.toLowerCase().includes(q)?'block':'none'}} |
| </script></body></html> |
