tinylm-instruct

TinyLM 275M — Instruct (SmolTalk SFT)

An instruction-tuned 275M small language model. This is the E3-full base (Run D — MLA + Muon — continued-pretrained on a 7.34B-token distill-mixture) supervised-fine-tuned for one epoch on SmolTalk with a ChatML-style template and prompt-loss masking.

• Source code: https://github.com/shivnarainms22/TinyLM
• Base model: Shiv-22/tinylm-checkpoints-v2 → e3_full/step_06999.pt — the final E3-full checkpoint from the v2 continued-pretraining track. Note this is not Shiv-22/tinylm (Run D); E3-full is Run D continued on a 7.34B-token distill mixture (see results/v2/E3full_vs_runD.md)
• SFT details & eval: results/v3/sft_vs_base.md
• 📄 What the project found (all four tracks, one page): results/FINDINGS.md

TinyLM model family

Source & full results: github.com/shivnarainms22/TinyLM

Scope. This is a 275M research model, part of a training-methods portfolio. It follows instructions and holds a chat format, but it is not a reliable assistant — expect factual slips, weak multi-step reasoning, and occasional repetition. See Limitations.

---

Prompt format

The model was trained on this exact text template (no new special tokens; vocab stays 32,000). Encode without a BOS token; the model closes its own turn with EOS.

<|user|>
{your message}
<|assistant|>

An optional system turn may precede it: <|system|>\n{system}\n. Generation continues after the final <|assistant|>\n and stops at EOS. The repo's scripts/generate_sft_samples.py builds this priming string exactly.

Usage

This is a custom PyTorch model (not a transformers architecture), so load it with the repo's code rather than AutoModelForCausalLM:

git clone https://github.com/shivnarainms22/TinyLM && cd TinyLM
pip install torch transformers huggingface_hub

from huggingface_hub import hf_hub_download
from transformers import AutoTokenizer
import sys; sys.path.insert(0, "src")
from tinylm.loader import load_from_checkpoint
from tinylm.sft import render_chat, _ROLE_PREFIX

ckpt = hf_hub_download("Shiv-22/tinylm-instruct", "tinylm_sft_smoltalk.pt")
model = load_from_checkpoint(ckpt, device="cpu").eval()

tok = AutoTokenizer.from_pretrained("meta-llama/Llama-2-7b-hf")
enc = lambda s: tok.encode(s, add_special_tokens=False)  # no BOS
ids, _ = render_chat([{"role": "user", "content": "Explain photosynthesis simply."}],
                     enc, tok.eos_token_id)
ids += enc(_ROLE_PREFIX["assistant"])   # prime the assistant turn
# ... greedy-decode from `ids`, stop at eos — see scripts/generate_sft_samples.py

scripts/generate_sft_samples.py --checkpoint <path> does exactly this end to end.

Training

Training logs: SFT loss curve on Weights & Biases — tinylm-v3. The base model's continued-pretraining runs are in tinylm-v2.

Evaluation — the alignment tax

Instruction-tuning is evaluated honestly for its cost on the locked completion/MCQ suite (it cannot measure instruction-following, only what SFT trades away). Base = E3-full at 0-shot; both scored with the same harness.

Language modeling is preserved (LAMBADA perplexity even nudges down — no catastrophic forgetting), commonsense is flat, and the one real cost is ~2.5 points on ARC-Easy multiple-choice — the expected tax of shifting the model from ranking canned options toward producing chat answers.

Sample outputs

Greedy decoding, unedited. Full set (8 prompts): results/v3/sft_samples.md.

Explain photosynthesis in one simple paragraph.

Photosynthesis is the process by which plants convert light energy into chemical energy, which is used to produce food for the plant. The process involves the absorption of light energy by chlorophyll in the leaves of a plant, which then absorbs the energy and uses it to produce glucose, a type of sugar. […]

Rewrite this sentence to be more polite: 'Send me the file now.'

I'll send you the file now.

Give me two tips for staying focused while studying.

First, establish a study routine that works for you. Set aside dedicated time each day to study… Another key tip is to create a dedicated study space… turn off any distractions like phones or TVs.

The model reliably adopts the assistant role, stays on topic, and produces coherent, on-format answers — that is the instruction-following the benchmark table above cannot see.

Limitations

These are real and visible in the same sample set — shown, not hidden:

• Weak multi-step reasoning / arithmetic. Asked a train's average speed (60 km / 2 h), it answered "120 km/h" then "≈1.25" — wrong. Consistent with the v2/v3 finding that reasoning at 275M is capacity-bound, and SFT does not add reasoning the base model lacks.
• Factual errors. "good morning" → Spanish came back as "¿Dónde está el día?" (incorrect). It will state wrong facts confidently.
• Repetition on open-ended generation. Under greedy decoding, creative prompts (e.g. a poem) can collapse into a repeated line. Sampling (temperature 0.7, top_p 0.9) mitigates this but trades determinism.
• No safety tuning. Only SmolTalk SFT — no preference optimization (DPO/RLHF), no safety alignment. Do not deploy in user-facing or high-stakes settings.

Intended use

Research and education: studying small-model post-training, prompt-template effects, and the gap between benchmark scores and instruction-following. Not intended for production assistance.