Base model and use
- Base:
Qwen/Qwen3-14B, revision
40c069824f4251a91eefaf281ebe4c544efd3e18
- Framework: PEFT 0.20.0
- Task: return speaker-attribution JSON under Alexandria's attribution prompt
- LoRA: rank 8, alpha 16, dropout 0.05
- Target modules: q/k/v/o and gate/up/down projections
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
base_id = "Qwen/Qwen3-14B"
tokenizer = AutoTokenizer.from_pretrained(base_id)
base = AutoModelForCausalLM.from_pretrained(base_id, device_map="auto")
model = PeftModel.from_pretrained(
base, "Om22s/alexandria-qwen3-14b-speaker-attribution-r8-s20260904"
)
The adapter was trained against Alexandria's structured attribution format;
ordinary chat prompts are outside the measured use case.
Training provenance
The unmodified training_manifest.json records the training code commit,
harness hash, data hash, seeds, and recipe. The source machine used a local
snapshot path, so adapter_config.json was made portable by replacing only
base_model_name_or_path with Qwen/Qwen3-14B. The original generated config
is retained as adapter_config.training-original.json.
- Training/data seed: 20260904
- Epochs: 1
- Learning rate: 2e-5
- Precision: BF16 LoRA (not 4-bit QLoRA)
- Training mixture SHA-256:
ab1bc2565444e8a96f11ed2835c1d5e70741a64bad1226de46126ca25e7ad8ba
The training mixture itself is not included. Its filename alone is not enough
to establish the licensing or composition of every source row, so this card
does not make a stronger dataset claim than the preserved manifest supports.
Evaluation
Paired base-versus-adapter evaluation used identical prompts, deterministic
generation, and 383 difficult gold-labelled dialogue rows across three light
novels. These are project-created hard-subset fixtures, not representative
whole-book samples.
Table with columns: book, rows, base, adapter, delta| book | rows | base | adapter | delta |
|---|
| Index 18 | 88 | 61.4% | 73.9% | +12.5 |
| Mushoku Tensei 16 | 133 | 51.9% | 54.1% | +2.3 |
| Owarimonogatari 3 | 162 | 38.9% | 35.2% | -3.7 |
|
The mixed per-book result is why this adapter is labeled experimental. It
improved two books but regressed the largest one. evaluation.json contains
all 766 arm rows, gold-file hashes, generation diagnostics, and provenance.
Limitations
- Evaluated on three translated light-novel hard subsets, not general prose.
- The pooled gain is modest and does not establish broad generalization.
- Speaker labels and output format are specific to Alexandria's prompt.
- This adapter does not contain the base-model weights.
Use remains subject to the Apache-2.0 license and the base model's terms.