README.md

---

license: apache-2.0
base_model: XHToken/Spark-X2.5-4B
language:
- el
- en
library_name: llama.cpp
pipeline_tag: text-generation
tags:
- gguf
- greek
- quantized
- tool-calling
---


<p align="center">
  <img src="https://huggingface.co/ApollonLabs/Heliactis-1-4B-GGUF/resolve/main/insignia.svg" width="280" alt="ApollonLabsAI — Towards the Sun">
</p>

# Heliactis-1-4B — GGUF

> **Heliactis** means **"sunbeam"**. It is the ancient-style form of the living Greek word
> *ηλιαχτίδα* (*ηλιακτίδα*), built from the ancient *hḗlios* ("sun") and *aktís* ("ray"). The Heliactis models are the rays; the larger **Helios**
> models are the sun itself. The lab is named after **Apollo** (*Apollon*), the Greek god of light.

Quantized builds of **Heliactis 1**, an Apollon Labs fine-tune of
[`XHToken/Spark-X2.5-4B`](https://huggingface.co/XHToken/Spark-X2.5-4B),
for `llama.cpp`, LM Studio, Ollama and anything else that reads GGUF.

**What it is for:** Greek that does not fall apart, code that stays intact, and
**tool calling that knows when to stay quiet.**

## Files

| File | Size | Gate |
|---|---|---|
| `Heliactis-1-4B-F16.gguf` | 7.66 GiB | passed |
| `Heliactis-1-4B-Q8_0.gguf` | 4.07 GiB | passed |
| `Heliactis-1-4B-Q6_K.gguf` | 3.15 GiB | passed (imatrix) |
| `Heliactis-1-4B-Q5_K_M.gguf` | 2.77 GiB | passed (imatrix) |
| `Heliactis-1-4B-Q4_K_M.gguf` | 2.42 GiB | passed (imatrix) · **the default**, runs on 6 GB VRAM or plain CPU |

Every file listed here went through the full gate below. Nothing is shipped unmeasured.

## Usage

```bash

llama-server -m Heliactis-1-4B-Q4_K_M.gguf --jinja -c 8192 \

             --cache-type-k q8_0 --cache-type-v q8_0

```

**Context window: 1,048,576 tokens**, inherited unchanged from the base
(`max_position_embeddings`; 27 of 36 layers use a 512-token sliding window, which keeps
the KV cache small). `-c 8192` above is only a default that fits small GPUs: raise it as
far as your memory allows. The KV cache at `q8_0` costs 14,976 bytes per token, about
14.6 GiB for the full window. Long-context recall was **not re-measured** for this release;
see the limitations below.

The chat template is embedded in the GGUF. The model was trained and evaluated
with **thinking disabled**. Call it that way. `--jinja` is required for tool
calling, otherwise the toolbox never reaches the model.

### Identity

This model was **not trained on any self-identity data**. If identity matters,
set it in the system prompt:

```

You are Heliactis 1, a Greek/English assistant built by Apollon Labs

on top of Spark-X2.5-4B.

```

## A base-model defect, reduced

On `Spark-X2.5-4B`, 502 vocabulary entries begin with an orphan UTF-8 byte. When the model
emits one, the output breaks and `llama-server` answers HTTP 500, even on an ordinary
translation prompt. We [reported it upstream](https://huggingface.co/XHToken/Spark-X2.5-4B/discussions/25).
Heliactis 1 was trained against those tokens: at the worst position we measured, their
probability fell from **12.7%** (our unreleased intermediate fine-tune, the starting point of
this training) to **0.17%**. Rarer, not impossible.

**Block them at inference** to close this route (`ban-ids.json` ships in this repo):

```bash

llama-server -m Heliactis-1-4B-Q4_K_M.gguf --jinja \

  $(python -c "import json;print(' '.join(f'-l {i}-inf' for i in json.load(open('ban-ids.json'))))")

```

Via the API, per request: `"logit_bias": [[<id>, false], ...]` for the same ids.
A system prompt does not fix it. Full reproduction, training trade-off and a second
route that blocking does not close: [DEFECT.md](DEFECT.md).

## Measured results

Instrument: 12-part gate, thinking disabled, `--jinja`, KV `q8_0`, `-c 8192`.
Each value comes from the gate's own results files.

| Test | Q4_K_M | Q5_K_M | Q6_K | Q8_0 | F16 | bar |
|---|---|---|---|---|---|---|
| Tool abstention | 15/15 | 15/15 | 15/15 | 15/15 | 15/15 | ≥ 9 |
| Overall tool score | 45/45 | 45/45 | 45/45 | 45/45 | 45/45 | ≥ 28 |
| Tool calls, Greek | 10/10 | 10/10 | 10/10 | 10/10 | 10/10 | ≥ 7 |
| Tool calls, English | 20/20 | 20/20 | 20/20 | 20/20 | 20/20 | ≥ 17.5 |
| HumanEval pass@1 | 0.7439 | 0.7622 | 0.7805 | 0.7622 | 0.7866 | ≥ 0.6952 |
| Code fencing | 60/60 | 57/60 | 58/60 | 59/60 | 58/60 | ≥ 53 |
| Capability eval | 52/60 | 53/60 | 55/60 | 54/60 | 54/60 | ≥ 51 |
| Greek character validity | 0.9996 | 0.9983 | 0.9974 | 0.9974 | 0.9968 | ≥ 0.9943 |
| Greek vocabulary | 0.9553 | 0.9536 | 0.9528 | 0.9549 | 0.9568 | ≥ 0.949 |
| Greek meaning | 0.76 | 0.79 | 0.78 | 0.79 | 0.80 | ≥ 0.70 |
| Median answer length (tokens) | 242 | 251.5 | 256.5 | 253.5 | 254.5 | ≤ 266 |
| Truncated answers | 3 | 5 | 2 | 2 | 3 | ≤ 9 |


### Honest limitations

- **The orphan-byte defect is reduced, not removed.** See above; block the ids to close this route, and see the Lao note below.
- **Do not use this model for Lao.** A second route to the same broken output exists: the model
  can emit a token that ends mid-character and then not finish it. Blocking the 502 ids does not
  stop this. In our test (greedy, one prompt per language) a Lao paragraph ended in HTTP 500
  on this model and on our unreleased intermediate fine-tune, but not on the base model; Thai passed on all three.
- **Lao and Thai got worse.** The 502 blocked tokens are pieces of those scripts.
  Perplexity vs our unreleased intermediate fine-tune, fp32, 5 short outside texts (151 tokens,
  so the noise is large): English 17.09 → 16.81 · code 18.44 → 19.30 · Greek 21.13 → 27.13 ·
  mixed languages 56.73 → 67.90 · Lao/Thai 56.49 → 78.85. The Greek line is one sentence;
  the gate's Greek vocabulary test is the measure we trust, and it passed.
- **Greek meaning (0.76 on Q4_K_M)** has a wide ±0.08 band. It detects collapse, not small changes.
- **Abstention was measured on N=15 prompts.** Evidence, not proof.
- **Long context was not re-measured for this release.** The architecture and KV cost
  are the base model's (14,976 bytes/token at q8_0; 9 of 36 layers full attention).

  An unreleased earlier fine-tune on the same base recalled 3/3 two-hop facts at 231k tokens.

  We do not carry that number over as a claim for this model.

- **24 of the 554 code training examples use Greek identifiers.** Known, not yet fixed.



## Training



LoRA (r=16, alpha=32). Stage 1: SFT on 2014 examples (1212 Greek, 554 code,

248 tool-calling; 742 of the 866 toolbox rows are decoys that teach abstention).

Dataset sha256: `cd58f73ad7fbb339de87a3bef22af6c2ff996b971f829978c78b4f70703f26e6`.

Stage 2: on-policy unlikelihood on the 502 tokens (856 of the model's own answers,

weight 10), with the same SFT set alongside to hold the rest in place.

imatrix: 100 × 512-token chunks of Greek text.



## License and attribution



Apache 2.0, inherited from the base model `XHToken/Spark-X2.5-4B`.

Derivative work by Apollon Labs.



---

*Apollon Labs — Towards the Sun.*