7-8b-open-weight-instruct-models-fully-fine-tuned-single-domain-dataset CLM-0345When 7-8B open-weight instruct models are fully fine-tuned on a single domain dataset, confining gradient updates to FFN neurons that activate stably across the training set and adding an MSE loss that pulls activations back toward pre-computed activation anchors retains far more general-benchmark ability (MMLU, BBH, TyDiQA) than full-parameter fine-tuning, by about 9.9 points on average, while also raising in-domain scores.
Capability: Retaining skills and safety after fine-tuning
Observed on
Open-weight models with accessible internals (Llama-3.1-8B, Qwen2.5-7B, Mistral-7B-v0.3, plus Qwen2.5 variants 0.5B-14B); single-domain supervised fine-tuning on GSM8K, DialogSum o.
Sources
- supportsNeuron-Guided Fine-Tuning: Unlocking Efficient Alignment Mechanisms for Large Language ModelsValidated by: In-domain metric (GSM8K accuracy, ROUGE-L on DialogSum/BioInstruct) and a CF score averaging MMLU, BBH and TyDiQA, against full-parameter fine-tuning, a composite strong baseline (Data Whisperer + NCFT + CE), and per-module baselines (NeFT, NCFT, MathNeuro, LESS, Nuggets, CCS) at 1/5/10/100% data; reported gains of 4.66 in-domain and 9.89 CF over FPFT averaged over three models; ablations isolate neuron selection, data selection and the loss.. [truncated]
Disagreeing is the most useful thing you can do here. Nothing on this page waits for a reviewer: the counts above move when evidence lands, and every source stays attached with who brought it.
Related claims
- For small open-weight Qwen2.5 models (1.5B-3B) doing multi-step math reasoning, features derived from the spread of attention entropy across layers at each generation step predict whether the final answer is wrong better than output-logit confidence features, with the margin larger on harder problems (+0.0093 AUC on GSM8K, +0.0759 on MATH-500 at 3B).Stating false facts confidently · reviewed by AI
- For a domain-specific language essentially absent from pretraining corpora (FORM symbolic algebra), QLoRA fine-tuning an 8B model on a few thousand examples validated by executing them through the language's own interpreter beats frontier open-weight models of up to 756B parameters given syntax documentation in context, on execution rate and strict output match for the larger benchmarks, while general reasoning scores drop under 3 points.Generating and editing working code · reviewed by AI
- For 4B-8B open base models mid-trained for tool use, grounding-oriented supervision (documentation/code turned into QA) and execution-oriented supervision (executable trajectories from real APIs and MCP skills) help different things, and only the combination improves everything: trajectories alone give the bigger function-calling gain while grounding alone transfers better to interactive and unseen-tool benchmarks, and each branch alone scores below no mid-training on MCP-Universe.Using the tools it is given · reviewed by AI
- Across 25 open-weight instruction-tuned models (4B-744B), the ability to modulate a named concept's residual-stream projection on plain instruction does not track parameter count or recency — the best scorer is an 8B 2024 model and some of the newest models score lowest — while within a single model line the score rises steadily across ordinary pre-training and post-training checkpoints, so activation controllability is acquired incidentally and must be measured per model rather than inferred from scale.Strategic deception and detecting it · reviewed by AI
- When post-training a 7-8B open model for text-to-SQL with a fixed budget of synthetic examples drawn from a large external pool, selecting examples whose schema-independent SQL skeletons are farthest from those already in the seed set (greedy K-center) rather than sampling randomly from the same pool yields a small in-domain gain and a larger cross-domain one (BIRD dev 64.9 vs 64.4, Spider Test 85.3 vs 83.4, Science Benchmark 53.9 vs 49.8).Writing a reusable locator: regex, SQL, XPath · reviewed by AI
Notes
Drafted from the paper by a model and filed unreviewed. Visible here so it can be read, not because anyone has vouched for it: it does not move any technique's standing and does not count toward the internal scorecard. Drafted confidence: medium. Falsifier as drafted: Applying the same neuron masking and activation-anchoring loss on other models or domain datasets and finding general-benchmark scores no better than plain full-parameter fine-tuning, or finding the gain comes only from updating fewer parameters (a random-neuron mask of the same size doing as well).