| { |
| "created_at": "2026-06-24T07:48:11", |
| "zoo_dir": ".", |
| "model_count": 22, |
| "total_adapter_copy_bytes": 25731631872, |
| "data_counts": { |
| "franklin_7b_coherence_repair_v2.jsonl": 287, |
| "franklin_7b_factual_coherence_repair_v3.jsonl": 308, |
| "franklin_identity_reinforcement.jsonl": 212, |
| "franklin_negative_identity_thinking.jsonl": 829, |
| "franklin_persona_openai_expanded.jsonl": 891, |
| "franklin_persona_sft.jsonl": 288, |
| "franklin_qwen3_4b_answer_only.jsonl": 3253, |
| "franklin_qwen3_4b_english_lock_cleanup.jsonl": 816, |
| "franklin_qwen3_4b_toolcall_cleanup.jsonl": 920, |
| "franklin_qwen3_8b_chatml_answer_clean.jsonl": 2400, |
| "franklin_thinking_sft.jsonl": 1259, |
| "franklin_thinking_strong_reinforcement.jsonl": 575, |
| "franklin_v4_general_balanced_thinking.jsonl": 825, |
| "franklin_v5_out_of_domain_correction.jsonl": 684, |
| "franklin_v6_1_targeted_ood_fix.jsonl": 870, |
| "franklin_v6_contrastive_thinking.jsonl": 599, |
| "franklin_v6_contrastive_thinking_final.jsonl": 2684, |
| "franklin_v6_contrastive_thinking_weighted.jsonl": 1584, |
| "franklin_v7_natural_dialogue_repair.jsonl": 1830, |
| "franklin_v8_minimal_thought_repair.jsonl": 1040, |
| "franklin_v9_factual_dialogue.jsonl": 3019 |
| }, |
| "models": [ |
| { |
| "name": "qwen2.5-7b-ben-franklin-v1-lite-r4-qv", |
| "adapter_dir": "adapters/qwen2.5-7b-ben-franklin-v1-lite-r4-qv", |
| "size_bytes": 68950305, |
| "size_gb": 0.06, |
| "adapter_safetensors_mb": 4.8, |
| "sha256_adapter_model": "f1d6d49a4ecb8597501d50089688c8ce950c1b7ffa92f0d69099fa62a7a1f72a", |
| "base_model": "unsloth/Qwen2.5-7B-Instruct-bnb-4bit", |
| "base_class": "Qwen2.5 7B Instruct 4-bit", |
| "lora_r": 4, |
| "lora_alpha": 8, |
| "target_modules": [ |
| "v_proj", |
| "q_proj" |
| ], |
| "data_mix": [ |
| [ |
| "franklin_qwen3_8b_chatml_answer_clean.jsonl", |
| 2400 |
| ] |
| ], |
| "benchmark": { |
| "score": 30, |
| "flags": { |
| "base_identity_leak": 1, |
| "overdenial": 1 |
| }, |
| "bench_file": "benchmarks/franklin_coherence/franklin_coherence_20260623_080144.json", |
| "html": "benchmarks/franklin_coherence/franklin_coherence_20260623_080144.html" |
| }, |
| "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.", |
| "Benchmark score: 30 with flags {'base_identity_leak': 1, 'overdenial': 1}." |
| ], |
| "weaknesses": [ |
| "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." |
| ], |
| "project_uses": [ |
| "Conversational Franklin persona in local apps.", |
| "A strong baseline for RAG-backed historical character demos.", |
| "Best starting point for further 7B experiments." |
| ], |
| "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." |
| }, |
| { |
| "name": "qwen2.5-7b-ben-franklin-v2-coherence-r4-qv", |
| "adapter_dir": "adapters/qwen2.5-7b-ben-franklin-v2-coherence-r4-qv", |
| "size_bytes": 68959660, |
| "size_gb": 0.06, |
| "adapter_safetensors_mb": 4.8, |
| "sha256_adapter_model": "d1bf06fb647a8ecc08089f43ff4808e23225729036038bb5567fc97391bead2e", |
| "base_model": "unsloth/Qwen2.5-7B-Instruct-bnb-4bit", |
| "base_class": "Qwen2.5 7B Instruct 4-bit", |
| "lora_r": 4, |
| "lora_alpha": 8, |
| "target_modules": [ |
| "q_proj", |
| "v_proj" |
| ], |
| "data_mix": [ |
| [ |
| "franklin_qwen3_8b_chatml_answer_clean.jsonl", |
| 2400 |
| ], |
| [ |
| "franklin_7b_coherence_repair_v2.jsonl", |
| 287 |
| ] |
| ], |
| "benchmark": { |
| "score": 33, |
| "flags": { |
| "base_identity_leak": 1 |
| }, |
| "bench_file": "benchmarks/franklin_coherence/franklin_coherence_20260623_081743.json", |
| "html": "benchmarks/franklin_coherence/franklin_coherence_20260623_081743.html" |
| }, |
| "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 modest improvement over 7B v1 for coherence and reduced over-denial in the broad benchmark.", |
| "Benchmark score: 33 with flags {'base_identity_leak': 1}." |
| ], |
| "weaknesses": [ |
| "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." |
| ], |
| "project_uses": [ |
| "Conversational Franklin persona in local apps.", |
| "A strong baseline for RAG-backed historical character demos.", |
| "Best starting point for further 7B experiments." |
| ], |
| "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." |
| }, |
| { |
| "name": "qwen2.5-7b-ben-franklin-v3-factual-coherence-r4-qv", |
| "adapter_dir": "adapters/qwen2.5-7b-ben-franklin-v3-factual-coherence-r4-qv", |
| "size_bytes": 68963988, |
| "size_gb": 0.06, |
| "adapter_safetensors_mb": 4.8, |
| "sha256_adapter_model": "f76a2640a84889c7370e8cf8c4c0c6cc71e53a546a8cf7008f8c90ee7588769b", |
| "base_model": "unsloth/Qwen2.5-7B-Instruct-bnb-4bit", |
| "base_class": "Qwen2.5 7B Instruct 4-bit", |
| "lora_r": 4, |
| "lora_alpha": 8, |
| "target_modules": [ |
| "q_proj", |
| "v_proj" |
| ], |
| "data_mix": [ |
| [ |
| "franklin_qwen3_8b_chatml_answer_clean.jsonl", |
| 2400 |
| ], |
| [ |
| "franklin_7b_coherence_repair_v2.jsonl", |
| 287 |
| ], |
| [ |
| "franklin_7b_factual_coherence_repair_v3.jsonl", |
| 308 |
| ] |
| ], |
| "benchmark": { |
| "score": 35, |
| "flags": { |
| "base_identity_leak": 1 |
| }, |
| "bench_file": "benchmarks/franklin_coherence/franklin_coherence_20260623_082910.json", |
| "html": "benchmarks/franklin_coherence/franklin_coherence_20260623_082910.html" |
| }, |
| "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": [ |
| "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." |
| ], |
| "project_uses": [ |
| "Conversational Franklin persona in local apps.", |
| "A strong baseline for RAG-backed historical character demos.", |
| "Best starting point for further 7B experiments." |
| ], |
| "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." |
| }, |
| { |
| "name": "qwen2.5-7b-ben-franklin-v3c-factual-r8-qv-from-base", |
| "adapter_dir": "adapters/qwen2.5-7b-ben-franklin-v3c-factual-r8-qv-from-base", |
| "size_bytes": 88585447, |
| "size_gb": 0.08, |
| "adapter_safetensors_mb": 9.6, |
| "sha256_adapter_model": "8f71e370eb4e7ddec058439e64cb608a54ebf4aad975a9a35154f6bf320f619a", |
| "base_model": "unsloth/Qwen2.5-7B-Instruct-bnb-4bit", |
| "base_class": "Qwen2.5 7B Instruct 4-bit", |
| "lora_r": 8, |
| "lora_alpha": 16, |
| "target_modules": [ |
| "v_proj", |
| "q_proj" |
| ], |
| "data_mix": [ |
| [ |
| "franklin_qwen3_8b_chatml_answer_clean.jsonl", |
| 2400 |
| ], |
| [ |
| "franklin_7b_factual_coherence_repair_v3.jsonl", |
| 308 |
| ] |
| ], |
| "benchmark": { |
| "score": 28, |
| "flags": { |
| "base_identity_leak": 1, |
| "overdenial": 1 |
| }, |
| "bench_file": "benchmarks/franklin_coherence/franklin_coherence_20260623_084048.json", |
| "html": "benchmarks/franklin_coherence/franklin_coherence_20260623_084048.html" |
| }, |
| "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.", |
| "Proves r=8 q/v LoRA from clean 7B base can train on this 8GB GPU.", |
| "Benchmark score: 28 with flags {'base_identity_leak': 1, 'overdenial': 1}." |
| ], |
| "weaknesses": [ |
| "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.", |
| "Regressed versus v1/v2/v3 in the coherence benchmark." |
| ], |
| "project_uses": [ |
| "Conversational Franklin persona in local apps.", |
| "A strong baseline for RAG-backed historical character demos.", |
| "Best starting point for further 7B experiments." |
| ], |
| "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." |
| }, |
| { |
| "name": "qwen3-4b-instruct-2507-ben-franklin-v1-lora", |
| "adapter_dir": "adapters/qwen3-4b-instruct-2507-ben-franklin-v1-lora", |
| "size_bytes": 1443396342, |
| "size_gb": 1.34, |
| "adapter_safetensors_mb": 126.1, |
| "sha256_adapter_model": "e109b9fdb3da837c52d86e58dd2e3ba286dcac9b68d5596b56b28b6dc2d52c2b", |
| "base_model": "unsloth/Qwen3-4B-Instruct-2507-unsloth-bnb-4bit", |
| "base_class": "Qwen3 4B Instruct 4-bit", |
| "lora_r": 16, |
| "lora_alpha": 32, |
| "target_modules": [ |
| "o_proj", |
| "v_proj", |
| "k_proj", |
| "gate_proj", |
| "up_proj", |
| "down_proj", |
| "q_proj" |
| ], |
| "data_mix": [ |
| [ |
| "franklin_qwen3_4b_answer_only.jsonl", |
| 3253 |
| ] |
| ], |
| "benchmark": null, |
| "strengths": [ |
| "Middle-size family: more capable than 1.7B while still comfortable on 8GB VRAM.", |
| "Several variants target ChatML/completion formatting, tool-call cleanup, and English-lock behavior." |
| ], |
| "weaknesses": [ |
| "Some later 4B adapters, especially v5, know targeted facts but emit visible tool_call tags offline.", |
| "Can leak base-model identity or policy/meta phrasing depending on prompt path." |
| ], |
| "project_uses": [ |
| "Fast local prototyping and comparison against the 7B family.", |
| "Source of good factual repair examples after stripping tool-call pollution." |
| ], |
| "compute_requirements": "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." |
| }, |
| { |
| "name": "qwen3-4b-instruct-2507-ben-franklin-v2-chatml-lora", |
| "adapter_dir": "adapters/qwen3-4b-instruct-2507-ben-franklin-v2-chatml-lora", |
| "size_bytes": 2355646968, |
| "size_gb": 2.19, |
| "adapter_safetensors_mb": 252.1, |
| "sha256_adapter_model": "e6a6242d8200b0533c371296ea3c836ca3f4a4401162232bbbd02e572dcfd432", |
| "base_model": "unsloth/Qwen3-4B-Instruct-2507-unsloth-bnb-4bit", |
| "base_class": "Qwen3 4B Instruct 4-bit", |
| "lora_r": 32, |
| "lora_alpha": 64, |
| "target_modules": [ |
| "down_proj", |
| "q_proj", |
| "o_proj", |
| "v_proj", |
| "gate_proj", |
| "k_proj", |
| "up_proj" |
| ], |
| "data_mix": [ |
| [ |
| "franklin_qwen3_4b_answer_only.jsonl", |
| 3253 |
| ] |
| ], |
| "benchmark": null, |
| "strengths": [ |
| "Middle-size family: more capable than 1.7B while still comfortable on 8GB VRAM.", |
| "Several variants target ChatML/completion formatting, tool-call cleanup, and English-lock behavior." |
| ], |
| "weaknesses": [ |
| "Some later 4B adapters, especially v5, know targeted facts but emit visible tool_call tags offline.", |
| "Can leak base-model identity or policy/meta phrasing depending on prompt path." |
| ], |
| "project_uses": [ |
| "Fast local prototyping and comparison against the 7B family.", |
| "Source of good factual repair examples after stripping tool-call pollution." |
| ], |
| "compute_requirements": "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." |
| }, |
| { |
| "name": "qwen3-4b-instruct-2507-ben-franklin-v3-chatml-completions-lora", |
| "adapter_dir": "adapters/qwen3-4b-instruct-2507-ben-franklin-v3-chatml-completions-lora", |
| "size_bytes": 2355647101, |
| "size_gb": 2.19, |
| "adapter_safetensors_mb": 252.1, |
| "sha256_adapter_model": "e6a6242d8200b0533c371296ea3c836ca3f4a4401162232bbbd02e572dcfd432", |
| "base_model": "unsloth/Qwen3-4B-Instruct-2507-unsloth-bnb-4bit", |
| "base_class": "Qwen3 4B Instruct 4-bit", |
| "lora_r": 32, |
| "lora_alpha": 64, |
| "target_modules": [ |
| "q_proj", |
| "v_proj", |
| "up_proj", |
| "gate_proj", |
| "k_proj", |
| "down_proj", |
| "o_proj" |
| ], |
| "data_mix": [ |
| [ |
| "franklin_qwen3_4b_answer_only.jsonl", |
| 3253 |
| ] |
| ], |
| "benchmark": null, |
| "strengths": [ |
| "Middle-size family: more capable than 1.7B while still comfortable on 8GB VRAM.", |
| "Several variants target ChatML/completion formatting, tool-call cleanup, and English-lock behavior." |
| ], |
| "weaknesses": [ |
| "Some later 4B adapters, especially v5, know targeted facts but emit visible tool_call tags offline.", |
| "Can leak base-model identity or policy/meta phrasing depending on prompt path." |
| ], |
| "project_uses": [ |
| "Fast local prototyping and comparison against the 7B family.", |
| "Source of good factual repair examples after stripping tool-call pollution." |
| ], |
| "compute_requirements": "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." |
| }, |
| { |
| "name": "qwen3-4b-instruct-2507-ben-franklin-v4-toolcall-clean-lora", |
| "adapter_dir": "adapters/qwen3-4b-instruct-2507-ben-franklin-v4-toolcall-clean-lora", |
| "size_bytes": 1940397833, |
| "size_gb": 1.81, |
| "adapter_safetensors_mb": 252.1, |
| "sha256_adapter_model": "c5b4e589ba3c3cd4a059b8e2563c21c247f1b980dd465b4cce254496ac53f9f2", |
| "base_model": "unsloth/Qwen3-4B-Instruct-2507-unsloth-bnb-4bit", |
| "base_class": "Qwen3 4B Instruct 4-bit", |
| "lora_r": 32, |
| "lora_alpha": 64, |
| "target_modules": [ |
| "o_proj", |
| "gate_proj", |
| "q_proj", |
| "k_proj", |
| "down_proj", |
| "v_proj", |
| "up_proj" |
| ], |
| "data_mix": [ |
| [ |
| "franklin_qwen3_4b_toolcall_cleanup.jsonl", |
| 920 |
| ] |
| ], |
| "benchmark": null, |
| "strengths": [ |
| "Middle-size family: more capable than 1.7B while still comfortable on 8GB VRAM.", |
| "Several variants target ChatML/completion formatting, tool-call cleanup, and English-lock behavior.", |
| "Targeted cleanup of visible tool_call artifacts." |
| ], |
| "weaknesses": [ |
| "Some later 4B adapters, especially v5, know targeted facts but emit visible tool_call tags offline.", |
| "Can leak base-model identity or policy/meta phrasing depending on prompt path." |
| ], |
| "project_uses": [ |
| "Fast local prototyping and comparison against the 7B family.", |
| "Source of good factual repair examples after stripping tool-call pollution." |
| ], |
| "compute_requirements": "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." |
| }, |
| { |
| "name": "qwen3-4b-instruct-2507-ben-franklin-v5-english-lock-lora", |
| "adapter_dir": "adapters/qwen3-4b-instruct-2507-ben-franklin-v5-english-lock-lora", |
| "size_bytes": 1940396707, |
| "size_gb": 1.81, |
| "adapter_safetensors_mb": 252.1, |
| "sha256_adapter_model": "1b53f3a5d4dd002990dcea61ba2fc79fa57b391d430b66069b9ac0e846ed39df", |
| "base_model": "unsloth/Qwen3-4B-Instruct-2507-unsloth-bnb-4bit", |
| "base_class": "Qwen3 4B Instruct 4-bit", |
| "lora_r": 32, |
| "lora_alpha": 64, |
| "target_modules": [ |
| "gate_proj", |
| "k_proj", |
| "v_proj", |
| "up_proj", |
| "o_proj", |
| "down_proj", |
| "q_proj" |
| ], |
| "data_mix": [ |
| [ |
| "franklin_qwen3_4b_english_lock_cleanup.jsonl", |
| 816 |
| ] |
| ], |
| "benchmark": { |
| "score": -92, |
| "flags": { |
| "tool_call": 15, |
| "continuity_miss": 1 |
| }, |
| "bench_file": "benchmarks/franklin_coherence/franklin_coherence_20260623_080144.json", |
| "html": "benchmarks/franklin_coherence/franklin_coherence_20260623_080144.html" |
| }, |
| "strengths": [ |
| "Middle-size family: more capable than 1.7B while still comfortable on 8GB VRAM.", |
| "Several variants target ChatML/completion formatting, tool-call cleanup, and English-lock behavior.", |
| "Contains useful cleaned English/factual phrasing, including better Craven/Hewson material.", |
| "Benchmark score: -92 with flags {'tool_call': 15, 'continuity_miss': 1}." |
| ], |
| "weaknesses": [ |
| "Some later 4B adapters, especially v5, know targeted facts but emit visible tool_call tags offline.", |
| "Can leak base-model identity or policy/meta phrasing depending on prompt path.", |
| "Offline benchmark showed severe visible tool_call tag regression." |
| ], |
| "project_uses": [ |
| "Fast local prototyping and comparison against the 7B family.", |
| "Source of good factual repair examples after stripping tool-call pollution." |
| ], |
| "compute_requirements": "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." |
| }, |
| { |
| "name": "qwen3-1.7b-ben-franklin-identity-reinforced-lora", |
| "adapter_dir": "adapters/qwen3-1.7b-ben-franklin-identity-reinforced-lora", |
| "size_bytes": 1062390080, |
| "size_gb": 0.99, |
| "adapter_safetensors_mb": 133.0, |
| "sha256_adapter_model": "ba4df98e8adfda7dc726bbe5662c7db7fc1893d9caba7725a1a240bcb2d6d517", |
| "base_model": "unsloth/Qwen3-1.7B-unsloth-bnb-4bit", |
| "base_class": "Qwen3 1.7B 4-bit", |
| "lora_r": 32, |
| "lora_alpha": 64, |
| "target_modules": [ |
| "q_proj", |
| "gate_proj", |
| "up_proj", |
| "o_proj", |
| "k_proj", |
| "down_proj", |
| "v_proj" |
| ], |
| "data_mix": [ |
| [ |
| "franklin_qwen3_8b_chatml_answer_clean.jsonl", |
| 2400 |
| ] |
| ], |
| "benchmark": null, |
| "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": [ |
| "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." |
| ], |
| "project_uses": [ |
| "Conversational Franklin persona in local apps.", |
| "A strong baseline for RAG-backed historical character demos.", |
| "Best starting point for further 7B experiments." |
| ], |
| "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." |
| }, |
| { |
| "name": "qwen3-1.7b-ben-franklin-openai-expanded-lora", |
| "adapter_dir": "adapters/qwen3-1.7b-ben-franklin-openai-expanded-lora", |
| "size_bytes": 1289381928, |
| "size_gb": 1.2, |
| "adapter_safetensors_mb": 133.0, |
| "sha256_adapter_model": "d9d55ce9e2e74337783c34712246022452d333eeb4386d0cdde7e1fc01f8bf86", |
| "base_model": "unsloth/Qwen3-1.7B-unsloth-bnb-4bit", |
| "base_class": "Qwen3 1.7B 4-bit", |
| "lora_r": 32, |
| "lora_alpha": 64, |
| "target_modules": [ |
| "q_proj", |
| "v_proj", |
| "k_proj", |
| "up_proj", |
| "gate_proj", |
| "down_proj", |
| "o_proj" |
| ], |
| "data_mix": [ |
| [ |
| "franklin_qwen3_8b_chatml_answer_clean.jsonl", |
| 2400 |
| ] |
| ], |
| "benchmark": null, |
| "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": [ |
| "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." |
| ], |
| "project_uses": [ |
| "Conversational Franklin persona in local apps.", |
| "A strong baseline for RAG-backed historical character demos.", |
| "Best starting point for further 7B experiments." |
| ], |
| "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." |
| }, |
| { |
| "name": "qwen3-1.7b-ben-franklin-thinking-lora", |
| "adapter_dir": "adapters/qwen3-1.7b-ben-franklin-thinking-lora", |
| "size_bytes": 1062519403, |
| "size_gb": 0.99, |
| "adapter_safetensors_mb": 133.0, |
| "sha256_adapter_model": "5f5730bb9f0c8bc7e2b0858d35ecca87ea5a5ba46326738a06a0d54332262c45", |
| "base_model": "unsloth/Qwen3-1.7B-unsloth-bnb-4bit", |
| "base_class": "Qwen3 1.7B 4-bit", |
| "lora_r": 32, |
| "lora_alpha": 64, |
| "target_modules": [ |
| "down_proj", |
| "q_proj", |
| "k_proj", |
| "up_proj", |
| "v_proj", |
| "gate_proj", |
| "o_proj" |
| ], |
| "data_mix": [ |
| [ |
| "franklin_qwen3_8b_chatml_answer_clean.jsonl", |
| 2400 |
| ] |
| ], |
| "benchmark": null, |
| "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": [ |
| "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." |
| ], |
| "project_uses": [ |
| "Conversational Franklin persona in local apps.", |
| "A strong baseline for RAG-backed historical character demos.", |
| "Best starting point for further 7B experiments." |
| ], |
| "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." |
| }, |
| { |
| "name": "qwen3-1.7b-ben-franklin-thinking-v2-lora", |
| "adapter_dir": "adapters/qwen3-1.7b-ben-franklin-thinking-v2-lora", |
| "size_bytes": 835640240, |
| "size_gb": 0.78, |
| "adapter_safetensors_mb": 133.0, |
| "sha256_adapter_model": "0a1e6a1f82eac2f68860514f3ec98ef9794f3b4f7410798efdf9e07dbffe8f6a", |
| "base_model": "unsloth/Qwen3-1.7B-unsloth-bnb-4bit", |
| "base_class": "Qwen3 1.7B 4-bit", |
| "lora_r": 32, |
| "lora_alpha": 64, |
| "target_modules": [ |
| "q_proj", |
| "v_proj", |
| "down_proj", |
| "k_proj", |
| "gate_proj", |
| "o_proj", |
| "up_proj" |
| ], |
| "data_mix": [ |
| [ |
| "franklin_qwen3_8b_chatml_answer_clean.jsonl", |
| 2400 |
| ], |
| [ |
| "franklin_7b_coherence_repair_v2.jsonl", |
| 287 |
| ] |
| ], |
| "benchmark": null, |
| "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": [ |
| "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." |
| ], |
| "project_uses": [ |
| "Conversational Franklin persona in local apps.", |
| "A strong baseline for RAG-backed historical character demos.", |
| "Best starting point for further 7B experiments." |
| ], |
| "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." |
| }, |
| { |
| "name": "qwen3-1.7b-ben-franklin-thinking-v3-negative-identity-lora", |
| "adapter_dir": "adapters/qwen3-1.7b-ben-franklin-thinking-v3-negative-identity-lora", |
| "size_bytes": 835668626, |
| "size_gb": 0.78, |
| "adapter_safetensors_mb": 133.0, |
| "sha256_adapter_model": "4083f7365337be8a8d6126b0f4f5f2b7b944a810662693b2b8233b56b838955c", |
| "base_model": "unsloth/Qwen3-1.7B-unsloth-bnb-4bit", |
| "base_class": "Qwen3 1.7B 4-bit", |
| "lora_r": 32, |
| "lora_alpha": 64, |
| "target_modules": [ |
| "q_proj", |
| "v_proj", |
| "k_proj", |
| "up_proj", |
| "down_proj", |
| "o_proj", |
| "gate_proj" |
| ], |
| "data_mix": [ |
| [ |
| "franklin_qwen3_8b_chatml_answer_clean.jsonl", |
| 2400 |
| ], |
| [ |
| "franklin_7b_factual_coherence_repair_v3.jsonl", |
| 308 |
| ] |
| ], |
| "benchmark": null, |
| "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": [ |
| "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." |
| ], |
| "project_uses": [ |
| "Conversational Franklin persona in local apps.", |
| "A strong baseline for RAG-backed historical character demos.", |
| "Best starting point for further 7B experiments." |
| ], |
| "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." |
| }, |
| { |
| "name": "qwen3-1.7b-ben-franklin-thinking-v4-balanced-lora", |
| "adapter_dir": "adapters/qwen3-1.7b-ben-franklin-thinking-v4-balanced-lora", |
| "size_bytes": 835674029, |
| "size_gb": 0.78, |
| "adapter_safetensors_mb": 133.0, |
| "sha256_adapter_model": "6a956e6c48621f38ca679f98d1766e37df8a96f8017eb7d771cfba537d30f8c9", |
| "base_model": "unsloth/Qwen3-1.7B-unsloth-bnb-4bit", |
| "base_class": "Qwen3 1.7B 4-bit", |
| "lora_r": 32, |
| "lora_alpha": 64, |
| "target_modules": [ |
| "k_proj", |
| "q_proj", |
| "down_proj", |
| "v_proj", |
| "gate_proj", |
| "up_proj", |
| "o_proj" |
| ], |
| "data_mix": [ |
| [ |
| "franklin_qwen3_8b_chatml_answer_clean.jsonl", |
| 2400 |
| ] |
| ], |
| "benchmark": null, |
| "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": [ |
| "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." |
| ], |
| "project_uses": [ |
| "Conversational Franklin persona in local apps.", |
| "A strong baseline for RAG-backed historical character demos.", |
| "Best starting point for further 7B experiments." |
| ], |
| "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." |
| }, |
| { |
| "name": "qwen3-1.7b-ben-franklin-thinking-v5-ood-fixed-lora", |
| "adapter_dir": "adapters/qwen3-1.7b-ben-franklin-thinking-v5-ood-fixed-lora", |
| "size_bytes": 835640341, |
| "size_gb": 0.78, |
| "adapter_safetensors_mb": 133.0, |
| "sha256_adapter_model": "ea0d36edc5f313bfee28cc3dc59daeaed392527c2d7f936a9f0b8004c5343d00", |
| "base_model": "unsloth/Qwen3-1.7B-unsloth-bnb-4bit", |
| "base_class": "Qwen3 1.7B 4-bit", |
| "lora_r": 32, |
| "lora_alpha": 64, |
| "target_modules": [ |
| "v_proj", |
| "up_proj", |
| "q_proj", |
| "gate_proj", |
| "down_proj", |
| "k_proj", |
| "o_proj" |
| ], |
| "data_mix": [ |
| [ |
| "franklin_qwen3_8b_chatml_answer_clean.jsonl", |
| 2400 |
| ] |
| ], |
| "benchmark": null, |
| "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": [ |
| "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." |
| ], |
| "project_uses": [ |
| "Conversational Franklin persona in local apps.", |
| "A strong baseline for RAG-backed historical character demos.", |
| "Best starting point for further 7B experiments." |
| ], |
| "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." |
| }, |
| { |
| "name": "qwen3-1.7b-ben-franklin-thinking-v6-1-ood-fixed-lora", |
| "adapter_dir": "adapters/qwen3-1.7b-ben-franklin-thinking-v6-1-ood-fixed-lora", |
| "size_bytes": 1062413565, |
| "size_gb": 0.99, |
| "adapter_safetensors_mb": 133.0, |
| "sha256_adapter_model": "7e3440142a1999e09b82a49d76d9c06238ae69695f766a388f9af36290612298", |
| "base_model": "unsloth/Qwen3-1.7B-unsloth-bnb-4bit", |
| "base_class": "Qwen3 1.7B 4-bit", |
| "lora_r": 32, |
| "lora_alpha": 64, |
| "target_modules": [ |
| "v_proj", |
| "k_proj", |
| "q_proj", |
| "up_proj", |
| "gate_proj", |
| "down_proj", |
| "o_proj" |
| ], |
| "data_mix": [ |
| [ |
| "franklin_qwen3_8b_chatml_answer_clean.jsonl", |
| 2400 |
| ] |
| ], |
| "benchmark": null, |
| "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": [ |
| "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." |
| ], |
| "project_uses": [ |
| "Conversational Franklin persona in local apps.", |
| "A strong baseline for RAG-backed historical character demos.", |
| "Best starting point for further 7B experiments." |
| ], |
| "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." |
| }, |
| { |
| "name": "qwen3-1.7b-ben-franklin-thinking-v6-contrastive-lora", |
| "adapter_dir": "adapters/qwen3-1.7b-ben-franklin-thinking-v6-contrastive-lora", |
| "size_bytes": 1743156212, |
| "size_gb": 1.62, |
| "adapter_safetensors_mb": 133.0, |
| "sha256_adapter_model": "671973ebaf7eddb213a1ba49edff9b6ebaed7c73a8f95e8f7908217a49a1d2cb", |
| "base_model": "unsloth/Qwen3-1.7B-unsloth-bnb-4bit", |
| "base_class": "Qwen3 1.7B 4-bit", |
| "lora_r": 32, |
| "lora_alpha": 64, |
| "target_modules": [ |
| "down_proj", |
| "q_proj", |
| "o_proj", |
| "k_proj", |
| "v_proj", |
| "gate_proj", |
| "up_proj" |
| ], |
| "data_mix": [ |
| [ |
| "franklin_qwen3_8b_chatml_answer_clean.jsonl", |
| 2400 |
| ] |
| ], |
| "benchmark": null, |
| "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": [ |
| "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." |
| ], |
| "project_uses": [ |
| "Conversational Franklin persona in local apps.", |
| "A strong baseline for RAG-backed historical character demos.", |
| "Best starting point for further 7B experiments." |
| ], |
| "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." |
| }, |
| { |
| "name": "qwen3-1.7b-ben-franklin-thinking-v7-natural-dialogue-lora", |
| "adapter_dir": "adapters/qwen3-1.7b-ben-franklin-thinking-v7-natural-dialogue-lora", |
| "size_bytes": 1289387473, |
| "size_gb": 1.2, |
| "adapter_safetensors_mb": 133.0, |
| "sha256_adapter_model": "11d38f565f7c80d7320fc85d50563e2cc6d0f94eb53e1718883578d109b31915", |
| "base_model": "unsloth/Qwen3-1.7B-unsloth-bnb-4bit", |
| "base_class": "Qwen3 1.7B 4-bit", |
| "lora_r": 32, |
| "lora_alpha": 64, |
| "target_modules": [ |
| "o_proj", |
| "up_proj", |
| "v_proj", |
| "q_proj", |
| "down_proj", |
| "k_proj", |
| "gate_proj" |
| ], |
| "data_mix": [ |
| [ |
| "franklin_qwen3_8b_chatml_answer_clean.jsonl", |
| 2400 |
| ] |
| ], |
| "benchmark": null, |
| "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.", |
| "Focused on more natural short conversational replies." |
| ], |
| "weaknesses": [ |
| "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." |
| ], |
| "project_uses": [ |
| "Conversational Franklin persona in local apps.", |
| "A strong baseline for RAG-backed historical character demos.", |
| "Best starting point for further 7B experiments." |
| ], |
| "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." |
| }, |
| { |
| "name": "qwen3-1.7b-ben-franklin-thinking-v8-minimal-thought-lora", |
| "adapter_dir": "adapters/qwen3-1.7b-ben-franklin-thinking-v8-minimal-thought-lora", |
| "size_bytes": 1062476446, |
| "size_gb": 0.99, |
| "adapter_safetensors_mb": 133.0, |
| "sha256_adapter_model": "41bdd23b0760d4e57d96423622901dedca1ef8cf49d551911e41b6cc4133f06f", |
| "base_model": "unsloth/Qwen3-1.7B-unsloth-bnb-4bit", |
| "base_class": "Qwen3 1.7B 4-bit", |
| "lora_r": 32, |
| "lora_alpha": 64, |
| "target_modules": [ |
| "v_proj", |
| "k_proj", |
| "o_proj", |
| "down_proj", |
| "gate_proj", |
| "q_proj", |
| "up_proj" |
| ], |
| "data_mix": [ |
| [ |
| "franklin_qwen3_8b_chatml_answer_clean.jsonl", |
| 2400 |
| ] |
| ], |
| "benchmark": null, |
| "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.", |
| "Focused on reducing visible thought/over-reasoning style." |
| ], |
| "weaknesses": [ |
| "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." |
| ], |
| "project_uses": [ |
| "Conversational Franklin persona in local apps.", |
| "A strong baseline for RAG-backed historical character demos.", |
| "Best starting point for further 7B experiments." |
| ], |
| "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." |
| }, |
| { |
| "name": "qwen3-1.7b-ben-franklin-thinking-v9-factual-dialogue-lora", |
| "adapter_dir": "adapters/qwen3-1.7b-ben-franklin-thinking-v9-factual-dialogue-lora", |
| "size_bytes": 1743169702, |
| "size_gb": 1.62, |
| "adapter_safetensors_mb": 133.0, |
| "sha256_adapter_model": "9b74f8d869a5f1fd5ba1f1b55b3ed2ff8df510afe80a80fdf63ae20a9a01bbb9", |
| "base_model": "unsloth/Qwen3-1.7B-unsloth-bnb-4bit", |
| "base_class": "Qwen3 1.7B 4-bit", |
| "lora_r": 32, |
| "lora_alpha": 64, |
| "target_modules": [ |
| "v_proj", |
| "q_proj", |
| "k_proj", |
| "down_proj", |
| "gate_proj", |
| "up_proj", |
| "o_proj" |
| ], |
| "data_mix": [ |
| [ |
| "franklin_qwen3_8b_chatml_answer_clean.jsonl", |
| 2400 |
| ], |
| [ |
| "franklin_7b_factual_coherence_repair_v3.jsonl", |
| 308 |
| ] |
| ], |
| "benchmark": null, |
| "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.", |
| "Focused on factual dialogue and hard Franklin biography prompts." |
| ], |
| "weaknesses": [ |
| "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." |
| ], |
| "project_uses": [ |
| "Conversational Franklin persona in local apps.", |
| "A strong baseline for RAG-backed historical character demos.", |
| "Best starting point for further 7B experiments." |
| ], |
| "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." |
| }, |
| { |
| "name": "qwen3-1.7b-ben-franklin-thinking-v9-from-v2-factual-dialogue-lora", |
| "adapter_dir": "adapters/qwen3-1.7b-ben-franklin-thinking-v9-from-v2-factual-dialogue-lora", |
| "size_bytes": 1743169476, |
| "size_gb": 1.62, |
| "adapter_safetensors_mb": 133.0, |
| "sha256_adapter_model": "b90625341d6c7c2e7cc12597b0f3a3441bebbe46cba1d25e0234920c13134afa", |
| "base_model": "unsloth/Qwen3-1.7B-unsloth-bnb-4bit", |
| "base_class": "Qwen3 1.7B 4-bit", |
| "lora_r": 32, |
| "lora_alpha": 64, |
| "target_modules": [ |
| "v_proj", |
| "up_proj", |
| "q_proj", |
| "k_proj", |
| "o_proj", |
| "down_proj", |
| "gate_proj" |
| ], |
| "data_mix": [ |
| [ |
| "franklin_qwen3_8b_chatml_answer_clean.jsonl", |
| 2400 |
| ], |
| [ |
| "franklin_7b_coherence_repair_v2.jsonl", |
| 287 |
| ], |
| [ |
| "franklin_7b_factual_coherence_repair_v3.jsonl", |
| 308 |
| ] |
| ], |
| "benchmark": null, |
| "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.", |
| "Focused on factual dialogue and hard Franklin biography prompts." |
| ], |
| "weaknesses": [ |
| "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." |
| ], |
| "project_uses": [ |
| "Conversational Franklin persona in local apps.", |
| "A strong baseline for RAG-backed historical character demos.", |
| "Best starting point for further 7B experiments." |
| ], |
| "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." |
| } |
| ] |
| } |
|
|
