README.md

---
license: apache-2.0
pipeline_tag: text-generation
language: en
tags:
  - tiny
  - tiny-lm
  - tiny-model
  - slm
  - SLM
  - small-language-model
  - from-scratch
  - llama
datasets:
  - HuggingFaceFW/fineweb-edu
  - allenai/dclm-baseline
metrics:
  - perplexity
---

# LDT-10M

A 10.28M-parameter LLaMA-style language model, trained from scratch on FineWeb-Edu + DCLM.

**Requested by [DedeProGames](https://huggingface.co/DedeProGames) on the [model-requests board](https://huggingface.co/spaces/Compactbot/model-requests) (#12).**

## Architecture

| Parameter | Value |
|---|---|
| Params | **10,284,480** |
| Layers | 5 |
| d_model | 320 |
| Heads | 5 (MHA, GQA not used at this scale) |
| FFN dim | 896 (SwiGLU) |
| Vocab | 12,288 (gollem BPE) |
| Context | 512 |
| Embeddings | Tied (lm_head → tok.weight) |
| Norm | RMSNorm (eps 1e-5) |
| Attention | RoPE + causal SDPA |
| Dtype | float32 |

Standard LLaMA block: RMSNorm → MHA (RoPE) → residual → RMSNorm → SwiGLU FFN → residual.

## Training

| | |
|---|---|
| Steps | 4,000 |
| Batch size | 64 |
| LR | 3e-4 → 3e-5 (cosine decay) |
| Data | FineWeb-Edu (30.1M tok) + DCLM baseline (21.4M tok) = **~50.5M tokens** |
| Tokens/param | **~4.9** |
| Hardware | RTX 5090 (32 GB), GPU |
| Final val loss | **4.6020** (ppl 99.68) |

### ⚠️ Honest caveat: undertrained

DedeProGames requested **2.6B tokens**. This checkpoint is at **50.5M tokens** — a ~50× shortfall. The GPU was occupied by other work for most of the training window, and the CPU was over-subscribed.

At 4.9 tok/param, the model has learned grammar and surface fluency but not deep coherence. The val loss (4.602) is well below the 7.38 unigram floor, so it genuinely uses context — but the prose is semantically thin: wordy, repetitive, and it drifts off-topic mid-sentence.

This is a **first checkpoint**, not the final deliverable. Continued training toward the 2.6B budget is planned.

## Eval (40 samples, 8 prompts × 5 seeds, temp 0.8, top-k 40)

| Metric | Value |
|---|---|
| mean loop_frac | 0.042 |
| max loop_frac | 1.0 (one sample) |
| Degenerate? | **No** |
| Below unigram floor? | **Yes** (4.602 < 7.38) |

### Sample outputs

> "The sun is assembled by a new study because he is an associate of the study of the disease in the early years. I've been interested in having a very different study…"

> "Once upon a time, she is an attack. But he is not a good idea. But it is something that does not have a moment of his own life…"

> "The cat sat on the ground. The sunp is a piece of light and is not a good deal of tear. The new story of the MD's Ford…"

> "def hello(): I have a lot of the best. I'm not sure what happened to me. I'll be able to do anything, but I think it's been just a lot to say…"

Grammar is intact. No token loops, no broken tokens, no speaker tags. Semantically limited — expected at 4.9 tok/param.

## Usage

The model uses a custom `LDT` architecture (standard LLaMA block, no special tricks). The safetensors file contains 47 tensors with tied embeddings (lm_head is not stored separately; it shares `tok.weight`).

To load with a custom model class, you need a small LLaMA-style implementation matching the config above. The training script (`train_ldt10m_fixed.py`) contains the full architecture definition.

## What this is NOT

- Not a 2.6B-token model (that's the target; this is the 50.5M checkpoint)
- Not a general-purpose assistant (it's a raw LM, no instruction tuning)
- Not a replacement for anything larger — it's a research checkpoint in a from-scratch training run