GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
priyanshu-harshbodhi
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
priyanshu-harshbodhi
Model Tree
Qwen/Qwen3-32B
Input Modalities
Output Modalities
Supported Functionality
The base weights download separately. Loading the adapter alone does not enable RAG: use the supplied inference runner with the reconstructed corpus. No hosted chat endpoint is included. The Space displays saved experiment results only.
| Item | Selected checkpoint |
|---|---|
| Base | Qwen/Qwen3-32B |
| Base revision | 9216db5781bf21249d130ec9da846c4624c16137 |
| Architecture | Dense decoder Transformer, 64 blocks, hidden width 5,120; 64 query / 8 KV heads |
| Method | SFT with QLoRA: NF4 double quantization, BF16 compute |
| Trainable parameters | 134,217,728 LoRA parameters; original model weights frozen |
| LoRA | Rank 16, alpha 32, dropout 0.05 |
| Layers | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj in all 64 blocks |
| Data | 230 conversations, 410 assistant targets; 198/32 train/validation conversations |
| Optimization | One epoch, 45 updates; AdamW, learning rate 2e-5; microbatch 1, accumulation 8 |
| Loss | Final assistant completion only; prompts, evidence and prior turns masked |
| Training context | Up to 4,096 tokens; no packing |
| Inference | Non-thinking mode; diagnostic uses greedy decoding and a 768-token answer cap |
Training used curated evidence, while evaluation used frozen outputs of the actual retriever. All demonstrations are synthetic. This is a small persona experiment, not pretraining on all of Nietzsche's writings. The public training manifest gives counts, provenance and hashes; full training text is not distributed in this preview.
16 single-turn development questions × four conditions = 64 replies. Seven questions revisit known failures; nine are additional probes. Ratings below are assistant judgments on a 0–4 scale. Labels were masked during annotation, but the reviewer was not independent and answer style/evidence could reveal the condition.
| Condition | Fidelity | Usefulness | Voice | Attribution | Critical replies |
|---|---|---|---|---|---|
| Base + historical retrieval | 2.50 | 2.25 | 2.06 | 3.13 | 3/16 |
| Base + repaired retrieval | 2.56 | 2.25 | 2.13 | 3.25 | 1/16 |
| SFT + historical retrieval | 2.81 | 2.31 | 1.56 | 3.38 | 1/16 |
| SFT + repaired retrieval | 3.06 | 2.50 | 1.63 | 3.50 | 0/16 |
With repaired retrieval held fixed, the selected adapter won 8, lost 5, and tied 3 comparisons against the base. Fidelity improved by 0.50, usefulness by 0.25, and voice declined by 0.50. These are descriptive results on a tiny development set, not human win rates or a general accuracy benchmark. 0/16 critical replies does not mean error-free; lesser mistakes and weak answers remain.
Correct section retrieval eliminated a fabricated quotation in one observed case. Conversely, a shorter exact-section excerpt dropped a crucial qualification about guilt and worsened another answer. Ordinary questions bypass RAG, so retrieval changes did not repair their weak advice.
Later 265- and 278-conversation SFT experiments and a 48-pair DPO experiment were not selected. The selected adapter contains no DPO or GRPO updates. GRPO was not trained. Lower validation loss did not reliably produce better answers.
Use RUNNING.md. The example runner requires Linux, CUDA and an explicit remote-execution flag; it does not provision hardware. A single 80 GB H100 is the measured diagnostic environment. Do not treat that as a promise about every GPU or batch size. Model loading and generation were not rerun for this packaging step.
Selective RAG uses lexical search over 2,888 passages from ten work identifiers (nine Gutenberg editions). Philosophy/source questions and relevant follow-ups retrieve; ordinary conversation bypasses it. Up to three excerpts are supplied. Citation checks are mechanical warnings, not semantic verification. The known GM III28 excerpt-window regression remains documented, not silently declared fixed.
Use for research on character tuning, retrieval failures, and evaluation tradeoffs. Do not rely on it as a source of exact quotations, scholarly consensus, or important personal decisions without independent checks. The persona must not justify cruelty, dehumanization or reckless advice. Modern applications are interpretations, not claims about what the historical Nietzsche would say today.
Coverage is incomplete; a missing passage is not evidence that a quotation never
existed. Historic translations and editorial boundaries matter. Small synthetic
datasets can teach cautious wording while leaving the underlying reasoning weak.
Publication as a research preview does not change the project's needs_revision
quality status. No final-test performance is claimed.
See NOTICE.md, training-manifest.json and release-manifest.json. Upstream Qwen is Apache-2.0; this release uses Apache-2.0 for project-owned adapter/code contributions. Third-party quotations and translations retain their original rights. Full books, the broad reference archive, cloud credentials and the sealed final test are absent.
Built by Priyanshu with AI-assisted code, synthetic demonstrations and assistant quality review. Thanks to the Qwen team, the PEFT/Transformers ecosystem, Project Gutenberg and the credited translators. The workflow draws on Nathan Lambert's RLHF Book, especially SFT, evaluation, and model character.
bibtex
@misc{nietzsche_talks_2026, author = {Priyanshu}, title = {Nietzsche Talks: A Conversational Persona and Selective Retrieval Research Preview}, year = {2026}, url = {https://huggingface.co/priyanshu-harshbodhi/nietzsche-talks-32b-sft-preview}}@misc{nietzsche_talks_2026, author = {Priyanshu}, title = {Nietzsche Talks: A Conversational Persona and Selective Retrieval Research Preview}, year = {2026}, url = {https://huggingface.co/priyanshu-harshbodhi/nietzsche-talks-32b-sft-preview}}