Boris-1.3-75M-Instruct-GGUF

GGUF quantizations of KSP-NMAI/Boris-1.3-75M-Instruct for llama.cpp and compatible runtimes (llama-server, LM Studio, koboldcpp, Jan).

The original safetensors weights live in the base repo — use those for finetuning or for any PyTorch-based runtime. GGUF is inference-only.

Which file should I pick?

Use Q8_0, or F16 if you want the exact reference weights.

Boris-1.3-75M is a small model, and quantization behaves differently at this scale than it does for 7B+ models. Two things set a hard floor on file size:

• The token embedding is 576 × 50304 ≈ 29M weights — 38% of the model — and llama.cpp keeps it at Q8_0 (~31 MB) in every quantized file here. The output weights are tied to it, so that 31 MB is unavoidable.
• d_model is 576 and the QKV/FFN-up rows are 1728; neither is divisible by 256, which k-quant and i-quant blocks require. Those tensors fall back to Q4_0, so only ffn_down (2304 columns) is ever quantized at the nominal bit width.

The practical result: every file here is between 54 MB and 151 MB. Dropping from Q8_0 to IQ1_S saves you a few tens of megabytes while degrading output substantially, and the ternary files are broken outright. The aggressive quants are provided for completeness, not because they are a good trade.

Files

Perplexity is measured on a fixed 30 × 1024-token slice of the calibration corpus. It is useful for comparing these rows against each other and nothing else.

File	Quant	Size	PPL	Notes
Boris-1.3-75M-Instruct-F16.gguf	F16	151M	71.04	Reference. Unquantized conversion of the safetensors weights.
Boris-1.3-75M-Instruct-BF16.gguf	BF16	151M	71.04	Reference, bfloat16.
Boris-1.3-75M-Instruct-Q8_0.gguf	Q8_0	82M	71.13	Effectively lossless. Recommended.
Boris-1.3-75M-Instruct-Q6_K.gguf	Q6_K	78M	71.05	Near-lossless.
Boris-1.3-75M-Instruct-Q5_K_M.gguf	Q5_K_M	70M	71.22	Very good quality.
Boris-1.3-75M-Instruct-Q5_K_S.gguf	Q5_K_S	67M	71.24
Boris-1.3-75M-Instruct-Q5_1.gguf	Q5_1	68M	71.85
Boris-1.3-75M-Instruct-Q5_0.gguf	Q5_0	65M	71.93
Boris-1.3-75M-Instruct-Q4_K_M.gguf	Q4_K_M	67M	71.71	Standard 4-bit default for larger models.
Boris-1.3-75M-Instruct-Q4_K_S.gguf	Q4_K_S	64M	72.08
Boris-1.3-75M-Instruct-Q4_1.gguf	Q4_1	62M	73.46
Boris-1.3-75M-Instruct-Q4_0.gguf	Q4_0	59M	74.05
Boris-1.3-75M-Instruct-IQ4_NL.gguf	IQ4_NL	59M	73.56
Boris-1.3-75M-Instruct-IQ4_XS.gguf	IQ4_XS	59M	73.60
Boris-1.3-75M-Instruct-Q3_K_L.gguf	Q3_K_L	64M	71.82
Boris-1.3-75M-Instruct-Q3_K_M.gguf	Q3_K_M	61M	72.84
Boris-1.3-75M-Instruct-Q3_K_S.gguf	Q3_K_S	57M	74.75
Boris-1.3-75M-Instruct-IQ3_M.gguf	IQ3_M	59M	73.29
Boris-1.3-75M-Instruct-IQ3_S.gguf	IQ3_S	57M	73.89
Boris-1.3-75M-Instruct-IQ3_XS.gguf	IQ3_XS	57M	73.89
Boris-1.3-75M-Instruct-IQ3_XXS.gguf	IQ3_XXS	57M	75.57
Boris-1.3-75M-Instruct-Q2_K.gguf	Q2_K	57M	74.75
Boris-1.3-75M-Instruct-Q2_K_S.gguf	Q2_K_S	56M	78.03
Boris-1.3-75M-Instruct-IQ2_M.gguf	IQ2_M	56M	76.42
Boris-1.3-75M-Instruct-IQ2_S.gguf	IQ2_S	55M	78.52
Boris-1.3-75M-Instruct-IQ2_XS.gguf	IQ2_XS	55M	79.67
Boris-1.3-75M-Instruct-IQ2_XXS.gguf	IQ2_XXS	55M	83.35
Boris-1.3-75M-Instruct-IQ1_M.gguf	IQ1_M	54M	91.06	Still clean English, but loops and confabulates.
Boris-1.3-75M-Instruct-IQ1_S.gguf	IQ1_S	54M	97.65
Boris-1.3-75M-Instruct-TQ2_0.gguf	TQ2_0	55M	728.94	Ternary; grammatical but semantically broken. Curiosity only.
Boris-1.3-75M-Instruct-TQ1_0.gguf	TQ1_0	54M	728.94	Ternary; as above, and gives the same greedy output as TQ2_0.

All K-quant, IQ and TQ files were produced with an importance matrix calibrated on 300 chunks of held-out data drawn from the model's own pretraining mixture (60% fineweb-edu / 40% dclm) plus rendered instruction turns; the matrix itself is published here as boris-1.3-75m-instruct.imatrix. Q8_0 and the legacy Q4_0/Q4_1/Q5_0/Q5_1 are plain round-to-nearest. SHA256SUMS covers every file.

Converted and quantized with llama.cpp build b9013 (e48034dfc).

Usage

# straight from the Hub
llama-server -hf KSP-NMAI/Boris-1.3-75M-Instruct-GGUF:Q8_0 --jinja

# or a local file
llama-server -m Boris-1.3-75M-Instruct-Q8_0.gguf --jinja

The Alpaca chat template is embedded in every file, so --jinja applies the correct prompt format automatically. A system message is folded into the front of the following instruction, matching how the model was fine-tuned.

Prompt format

Below is an instruction that describes a task. Write a response that appropriately completes the request.

### Instruction:
{your instruction}

### Response:

The response begins with a leading space and ends with <|endoftext|> (token 0), which is also the stop token.

Verification

Every file in this repo was checked to load as arch=gpt2 with the chat template present, and to complete Alpaca prompts with output that decodes as strict UTF-8 containing zero U+FFFD replacement characters — no mojibake at any quant level, including the ternary files. The GGUF vocab was diffed against the source tokenizer.json across accented Latin, Cyrillic, CJK, Hangul, Arabic, Hebrew, emoji (including ZWJ sequences), zero-width and whitespace-run samples: identical token ids and byte-exact detokenization.

Limitations

This is a very small instruction-tuned model. It will produce text that is frequently inaccurate, inconsistent, or offensive, and has received no alignment or safety tuning beyond supervised fine-tuning. Do not rely on it for factual information or deploy it without supervision.

License

Apache 2.0. Copyright 2026 Joseph Jones. See the base repository for the full notice.