---
library_name: transformers
pipeline_tag: text-generation
language:
  - en
base_model:
  - atenareply/smollm2-135m-noval
  - HuggingFaceTB/SmolLM2-135M
datasets:
  - atenareply/noval-corp-sft-small
  - HuggingFaceTB/smol-smoltalk
tags:
  - smollm2
  - sft
  - instruction-tuning
  - chatml
license: apache-2.0
---

# SmolLM2-135M Noval — Instruct (SFT)

Instruction-tuned domain assistant: SFT of the 135M CPT model on grounded domain instruction pairs blended with general instructions.

## Overview

- **Stage**: SFT (TRL SFTTrainer, ChatML, assistant-only loss)
- **Lineage**: SmolLM2-135M → CPT (noval) → **SFT** (this model)
- **Method**: SFT from the CPT'd checkpoint. ChatML with a `{% generation %}` block + assistant-only loss; grounded instruction pairs (teacher-distilled, generator≠judge, verbatim-number grounding filter [[genie]](#citations)) blended 60/40 with smol-smoltalk [[collapse]](#citations).
- **Domain**: fictional — Orbital Mining Corporation (OMC) technical docs + Mars Express telemetry.

## Training

| | |
|---|---|
| Dataset | noval-corp-sft-small — train 1,388 / val 43 (chat `messages`) |
| LR / epochs | 2e-5 cosine, warmup 0.03, 3 epochs, max_len 2048, eff_batch 32 |

## Evaluation

| Metric | Value | Note |
|---|---|---|
| train_loss | **2.257** | healthy curve, no overfit |
| eval_loss | **1.946** | per-epoch 2.048 → 1.959 → 1.946 |
| token accuracy | **0.57** | 0.47 → 0.57 |

_Eval is a held-out, deterministic verifiable harness (synthetic tasks); baseline = the pre-SFT ParamΔ model. See `noval-corp/scripts/eval_agentic.py`._

## Intended use & limitations

Domain Q&A / summarization / code-explanation over OMC / Mars-Express, in chat (ChatML) format.

**Limitations:**
- 135M instruct is intrinsically limited.
- Instruction-following eval was handed off externally (no in-repo task metrics).
- Fictional domain; sparse one-off entities hallucination-prone.

## Innovations tested

- **Local-subscription teacher distillation** — grounded pairs generated by Claude Code (generator ≠ judge), with a rule-based verbatim-number grounding gate [[genie]](#citations).

## Usage

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

tok = AutoTokenizer.from_pretrained("atenareply/smollm2-135m-noval-instruct")
model = AutoModelForCausalLM.from_pretrained("atenareply/smollm2-135m-noval-instruct")
msgs = [{"role": "user", "content": "What does the Orbital Mining Corporation do?"}]
inputs = tok.apply_chat_template(msgs, add_generation_prompt=True, return_tensors="pt", return_dict=True)
print(tok.decode(model.generate(**inputs, max_new_tokens=256)[0, inputs["input_ids"].shape[1]:], skip_special_tokens=True))
```

## Citations

- GENIE (grounded synthesis) — Mirza et al., ICLR 2024. arXiv:2401.14367
- Model collapse (≥50% real data) — Shumailov et al., 2024. arXiv:2404.01413

---
_Card generated by `noval-corp/scripts/gen_model_cards.py` (standardized across the noval-corp model family)._