catalyst404/samsara-qwen1.5b

Samsara v8 — Qwen2.5-1.5B fine-tuned on The Tibetan Book of the Dead

LoRA fine-tune of Qwen/Qwen2.5-1.5B-Instruct (merged) on a 3,161-pair synthetic Q/A dataset built mechanically from the books Glossary of Key Terms, the six-sages prayer, and the modern Samsara app structured data.

Part of the Samsara project.

Eval

99.4% pass on a 308-case structured evaluation (severity-weighted 99.6%). Perfect or near-perfect on:

• the six realms (god / antigod / human / animal / anguished-spirit / hell)
• the three and five poisons
• the six sages and their colours, realms, and implements
• Tibetan / Sanskrit etymology for core terms
• hallucination-trap refusals (non-existent realms, nth poison, wrong colours)
• the Samsara apps modern reframings of each realm

See SCOREBOARD.md in the project repo for full version history.

Files

• onnx/model_q8.onnx — INT8-quantized ONNX (~1.7 GB) for Transformers.js (in-browser, WebGPU or WASM).
• tokenizer + config files for chat-template inference.

Load in the browser

import { pipeline } from "@huggingface/transformers";
const pipe = await pipeline("text-generation", "thealch3m1st/samsara-qwen1.5b", {
  dtype: "q8", device: "webgpu"
});
const out = await pipe([
  { role: "system", content: "You are a careful reader of The Tibetan Book of the Dead." },
  { role: "user",   content: "Who is the sage of the hell realms?" }
], { max_new_tokens: 220 });
console.log(out[0].generated_text.at(-1).content);

Known limitations

• 2/308 eval cases fail: (1) a stubborn "pride/god-realms vs heavenly-realms" linguistic quirk where the model answers "No — pride is for heavenly realms, envy for god realms" (internally self-contradicting); (2) Chapter 11 structural content (severity-1 nice-to-have).
• Training set is mechanical drills, not free paraphrase. For highly out-of-distribution phrasings or open-ended scholarly questions, fall back to RAG over the book corpus.
• Page-level citation text comes from RAG at inference time, not the models memory. Do not trust verbatim book quotes from this model alone.

Trained 2026-04-22 on DGX Spark (GB10, 128 GB unified) in 10 min.