Johnnyman1100

qwen2.5-1.5b-voynich-eva-lora

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizerfrom peft import PeftModel base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-1.5B-Instruct", dtype="bfloat16")tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-1.5B-Instruct")model = PeftModel.from_pretrained(base, "Johnnyman1100/qwen2.5-1.5b-voynich-eva-lora")

The included tokenizer.json, tokenizer_config.json and chat_template.jinja are Qwen's files, copied here for convenience.

What was actually measured

Training: 3 epochs, 237 optimizer steps, seq len 256, bf16, on 181 of 226 folios (45 held out; no multi-word line appears in both splits).

Held-out EVA perplexity: 22.32 (adapter) vs 133.42 (base Qwen)
Interpolated word-bigram baseline, same data: 1.66 nats/char vs adapter 1.31
vs base Qwen 2.07 — the adapter learns real EVA statistics beyond word adjacency

Generation controls (same prompt, seeds, settings, all 226 folios):

Feeding the adapter shuffled words or frequency-matched gibberish produces outputs within noise of real passages on all theme rates when pooled over the 226 folios (every difference under 8 points). Individual 45-folio held-out cells are noisier, and a few of them cross that threshold.
Its pseudo-Voynichese "echoes" occur at 36-45% on any letter-salad input — generic EVA vocabulary, not reading.
Removing the passage entirely is the only control that changes the distribution.

Bottom line: the adapter is a good statistical model of Voynichese. Its English "translations" are prompt-driven confabulation, independent of passage content.

Repository layout

adapter_model.safetensors, adapter_config.json — the LoRA adapter
scripts/ — training, generation, controls, baselines and analysis code
results/ — metrics, predictions (written before results), conclusions
outputs/ — raw per-folio JSONL for every run: adapter/base x seed42/43 x real/shuffled/gibberish/no-passage
story/ — The Scattered Leaves, a literary artifact composed from the adapter's speculative outputs, the editorial brief, the mined fragments, and a reading edition of all 226 folios. Fiction, explicitly.

Reproducing

The scripts were written for a workspace layout and default to paths that will not exist in a checkout of this repo. Every script accepts explicit paths — pass them:

bash

# 1. Get the source transcription (not redistributed here — see Provenance)#    Save it as ZL3b-n.txt next to the scripts.# 2. Trainpython scripts/finetune_qwen.py --model Qwen/Qwen2.5-1.5B-Instruct \    --source ZL3b-n.txt --output-dir ./my_run# 3. Generate for all folios (adapter, then base), plus controlspython scripts/translate_full_corpus.py --model ... --source ... --adapter ./my_run \    --output ./my_run/adapter.jsonl --prompts forced_speculationpython scripts/translate_full_corpus.py --no-adapter --control gibberish ...# 4. Baseline + analysispython scripts/baseline_bigram.py     # edit RUN_DIR / DEFAULT_SOURCE at toppython scripts/analyze_runs.py        # edit RUN_DIR / FILES at top

Split: split_folios(folios, 0.2, seed=42) — deterministic; the 45 held-out folio IDs are listed in results/run_metrics.json. Generation seeds are per-folio and derived from --seed, so a rerun with the same seed is reproducible modulo GPU nondeterminism.

The source/base_model/adapter fields inside the JSONL metadata records are the original local paths; they are informational only.

Provenance and licensing

Adapter weights, scripts, results, story: MIT (see LICENSE).
Base model: Qwen/Qwen2.5-1.5B-Instruct, Apache-2.0. Not included; downloaded automatically by PeftModel.from_pretrained. The included tokenizer.json, tokenizer_config.json, chat_template.jinja are Qwen's unmodified files (Apache-2.0).
Source text: ZL EVA 3b IVTFF transcription by René Zandbergen and Gabriel Landini, available from voynich.nu. Not redistributed here; it is their work. scripts/strip_ivtff.py documents exactly how it was cleaned.
Trained on a single RTX 3070 8GB, bf16, PEFT 0.18.1, Transformers 5.0.0, Torch 2.7.1.

Caution

Do not cite outputs of this adapter as evidence about the Voynich Manuscript's content, language, or provenance. The harness in scripts/ is designed to be reused to test other hypotheses the same way.

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Model Details

Model Provider

Johnnyman1100

Model Tree

Base

Qwen/Qwen2.5-1.5B-Instruct

Adapter

this model

Input Modalities

Text

Output Modalities

Text

Supported Functionality

Dedicated Endpoints

Explore FriendliAI today

Get started Talk to an engineer

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizerfrom peft import PeftModel base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-1.5B-Instruct", dtype="bfloat16")tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-1.5B-Instruct")model = PeftModel.from_pretrained(base, "Johnnyman1100/qwen2.5-1.5b-voynich-eva-lora")

from transformers import AutoModelForCausalLM, AutoTokenizerfrom peft import PeftModel base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-1.5B-Instruct", dtype="bfloat16")tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2.5-1.5B-Instruct")model = PeftModel.from_pretrained(base, "Johnnyman1100/qwen2.5-1.5b-voynich-eva-lora")

The included tokenizer.json, tokenizer_config.json and chat_template.jinja are Qwen's files, copied here for convenience.

What was actually measured

Training: 3 epochs, 237 optimizer steps, seq len 256, bf16, on 181 of 226 folios (45 held out; no multi-word line appears in both splits).

Held-out EVA perplexity: 22.32 (adapter) vs 133.42 (base Qwen)
Interpolated word-bigram baseline, same data: 1.66 nats/char vs adapter 1.31
vs base Qwen 2.07 — the adapter learns real EVA statistics beyond word adjacency

Generation controls (same prompt, seeds, settings, all 226 folios):

Feeding the adapter shuffled words or frequency-matched gibberish produces outputs within noise of real passages on all theme rates when pooled over the 226 folios (every difference under 8 points). Individual 45-folio held-out cells are noisier, and a few of them cross that threshold.
Its pseudo-Voynichese "echoes" occur at 36-45% on any letter-salad input — generic EVA vocabulary, not reading.
Removing the passage entirely is the only control that changes the distribution.

Bottom line: the adapter is a good statistical model of Voynichese. Its English "translations" are prompt-driven confabulation, independent of passage content.

Repository layout

adapter_model.safetensors, adapter_config.json — the LoRA adapter
scripts/ — training, generation, controls, baselines and analysis code
results/ — metrics, predictions (written before results), conclusions
outputs/ — raw per-folio JSONL for every run: adapter/base x seed42/43 x real/shuffled/gibberish/no-passage
story/ — The Scattered Leaves, a literary artifact composed from the adapter's speculative outputs, the editorial brief, the mined fragments, and a reading edition of all 226 folios. Fiction, explicitly.

Reproducing

The scripts were written for a workspace layout and default to paths that will not exist in a checkout of this repo. Every script accepts explicit paths — pass them:

bash

# 1. Get the source transcription (not redistributed here — see Provenance)#    Save it as ZL3b-n.txt next to the scripts.# 2. Trainpython scripts/finetune_qwen.py --model Qwen/Qwen2.5-1.5B-Instruct \    --source ZL3b-n.txt --output-dir ./my_run# 3. Generate for all folios (adapter, then base), plus controlspython scripts/translate_full_corpus.py --model ... --source ... --adapter ./my_run \    --output ./my_run/adapter.jsonl --prompts forced_speculationpython scripts/translate_full_corpus.py --no-adapter --control gibberish ...# 4. Baseline + analysispython scripts/baseline_bigram.py     # edit RUN_DIR / DEFAULT_SOURCE at toppython scripts/analyze_runs.py        # edit RUN_DIR / FILES at top

# 1. Get the source transcription (not redistributed here — see Provenance)#    Save it as ZL3b-n.txt next to the scripts.# 2. Trainpython scripts/finetune_qwen.py --model Qwen/Qwen2.5-1.5B-Instruct \    --source ZL3b-n.txt --output-dir ./my_run# 3. Generate for all folios (adapter, then base), plus controlspython scripts/translate_full_corpus.py --model ... --source ... --adapter ./my_run \    --output ./my_run/adapter.jsonl --prompts forced_speculationpython scripts/translate_full_corpus.py --no-adapter --control gibberish ...# 4. Baseline + analysispython scripts/baseline_bigram.py     # edit RUN_DIR / DEFAULT_SOURCE at toppython scripts/analyze_runs.py        # edit RUN_DIR / FILES at top

The source/base_model/adapter fields inside the JSONL metadata records are the original local paths; they are informational only.

Provenance and licensing

Adapter weights, scripts, results, story: MIT (see LICENSE).
Base model: Qwen/Qwen2.5-1.5B-Instruct, Apache-2.0. Not included; downloaded automatically by PeftModel.from_pretrained. The included tokenizer.json, tokenizer_config.json, chat_template.jinja are Qwen's unmodified files (Apache-2.0).
Source text: ZL EVA 3b IVTFF transcription by René Zandbergen and Gabriel Landini, available from voynich.nu. Not redistributed here; it is their work. scripts/strip_ivtff.py documents exactly how it was cleaned.
Trained on a single RTX 3070 8GB, bf16, PEFT 0.18.1, Transformers 5.0.0, Torch 2.7.1.

qwen2.5-1.5b-voynich-eva-lora

README

Usage

What was actually measured

Repository layout

Reproducing

Provenance and licensing

Caution

Explore FriendliAI today

README

Usage

What was actually measured

Repository layout

Reproducing

Provenance and licensing

Caution