TwIL-LM3

A 3B reasoning model for formal logic tasks, built from HuggingFaceTB/SmolLM3-3B through LoRA supervised fine-tuning, checkpoint fusion, WiSE-FT weight interpolation, and entropy-weighted GRPO reinforcement learning.

It improves in-domain formal-logic performance by +26% relative over its base model (macro gate 0.336 → 0.422) and improves held-out benchmark performance at the same time (+0.022 core average). It is the only arm in this project that gains on both tracks, which is why it is the recommended release of the pair.

Results

Track A — in-domain formal logic

All arms below were run through the same harness, prompts and decoding settings described under Evaluation protocol. Throughput rows are reported because in-domain score alone is misleading for a 3B model: ans/s is defined throughout as tok/s ÷ mean generation length, so it measures completed answers rather than raw decode rate.

lane / metric	TwIL-LM3	TwIL-LM3*	SmolLM3-3B base	Llama-3.2-3B	LFM2-2.6B	LFM2.5-8B-A1B	Qwen3-8B	gpt-oss-120b ‡
lean_formalize token_f1	0.5869	0.6456	0.4347	0.3690	0.1321	0.4655	0.4022	0.6306
rule_induction derivation	0.3192	0.9644	0.1029	0.0825	0.0615	0.1936	0.3680	0.6518
entailment_label accuracy	0.5750	0.6867	0.3750	0.3300	0.4700	0.5400	0.5800	0.7750
mcq_answer accuracy	0.1100	0.5200	0.0000	0.0000	0.0150	0.0750	0.0000	0.0700
semantic_parse token_f1	0.4416	0.8762	0.4149	0.3102	0.3665	0.3778	0.4257	0.4331
lean_critic accuracy	0.6600	0.5200	0.6500	0.5300	0.5900	0.5500	0.7950	0.5550
lm_corpus perplexity ↓	2.8972	3.1284	3.1818	2.8478	4.3815	4.9472	2.5440	912.23 §
math_corpus perplexity ↓	3.8229	3.5245	4.0685	4.7531	6.7472	8.3323	4.0083	1045.63 §
average, 6 lanes	0.4488	0.7021	0.3296	0.2703	0.2725	0.3670	0.4285	0.5192
macro gate	0.4218	0.5896	0.3466 †	0.2925	0.3473	0.3757	0.5336	—
strict-7	0.1971	0.3290	0.1493	0.1229	0.1579	0.1714	0.2093	—
macro_primary	0.4475	0.4958	0.4075	0.3450	0.4188	0.4213	0.5750	—
tok/s	15880	15840	15564	16160	25230	22480	9420	3374
mean gen length	564	572	999	696	2296	1830	2094	1005
ans/s	28.1	27.7	15.6	23.2	10.9	12.0	4.5	3.4

* TwIL-LM3* is our latest version of TwIL-LM3. The weights will be released soon — the files in this repository are the current TwIL-LM3 release, not this one. Lanes marked — are not yet reported for it.

‡ gpt-oss-120b runs MXFP4 weights at tensor-parallel 2 — quantized and multi-GPU, so its throughput rows are not directly comparable to the single-GPU BF16 arms. Its procedural lane and the loose-match scorings were not collected, so the three summary rows below the six-lane average cannot be computed for it; that is what the — cells mean, not a zero.

§ The 120B's perplexities are three orders of magnitude off every other arm because its harmony response format and tokenizer make the corpus lanes score a different quantity. The number is reported for completeness but is not a comparable measurement.

† The base column here comes from the external-comparison run rather than the paired base-vs-TwIL run, hence 0.3466 against the 0.3356 quoted in the summary at the top of this card — run-to-run variation of the same checkpoint. The paired run is the correct basis for the improvement claim.

average, 6 lanes is the plain mean of the six objective rows above it, each at whatever scoring that row reports. It is a coarser summary than the three that follow — it mixes token-F1 with accuracy — but it is the only summary row every arm here can be compared on, including the 120B.

The next three rows aggregate more carefully. None of them include the perplexity lanes or the token-F1 scorings, which are not on a common 0–1 accuracy scale.

macro gate is the headline metric and the one the training pipeline gates on. It is the equal-weight mean of five objectives: the four bounded classification lanes (entailment_label, mcq_answer, procedural, lean_critic) plus rule_induction, scored by its continuous derivation score. Rule induction is included specifically so a fine-tune cannot pass the gate while quietly regressing inductive reasoning. In the gate, mcq_answer and procedural are credited as max(exact_match, loose_match): for free-text answer lanes, a response that is correct but differently formatted is a formatting artefact rather than a reasoning failure. This affects the aggregate only — the per-lane rows above stay strict.

macro_primary is the same mean over the four classification lanes alone, without rule_induction. It is the narrower "bounded classification" view, kept for comparability with earlier reports; the gate is the metric to read for overall in-domain capability.

strict-7 is the mean of seven lanes scored under strict metrics only (fol_translation, entailment_label, mcq_answer, semantic_parse and lean_formalize exact match, lean_critic and procedural accuracy), with no loose-match credit anywhere. It is deliberately harsh — exact match on generative lanes is near zero for every arm — so it is useful for ranking models against each other but not as an absolute capability measure.

TwIL-LM3 beats every arm up to and including LFM2.5-8B-A1B, and does so on all six objective lanes and all four summary rows, not on average alone. Against the strongest of them it is 0.4218 to 0.3757 on the gate at roughly a third of the total parameters, with the margin coming from the lanes the pipeline targets directly: lean_formalize token-F1 0.5869 against 0.4655, rule_induction 0.3192 against 0.1936, semantic_parse 0.4416 against 0.3778.

It does not beat the two largest arms. Qwen3-8B leads it on the gate 0.5336 to 0.4218 and gpt-oss-120b leads the six-lane average 0.5192 to 0.4488. That gap is worth reading carefully in Qwen's case: almost all of it is loose-match credit. Qwen answers MCQ correctly but never in the requested format — strict accuracy 0.0000 against TwIL-LM3's 0.1100, while its loose match is 0.745 — and the macro rows credit max(exact_match, loose_match). On strict-7, which gives no loose-match credit anywhere, the two are 0.2093 to 0.1971, a gap of 0.012 rather than 0.11. Qwen also wins lean_critic outright at 0.7950 and has the lowest lm_corpus perplexity at 2.5440. The 120B leads three lanes outright and is genuinely stronger at entailment (0.7750) and rule induction (0.6518).

The size and speed context matters for both. Qwen3-8B is 2.6x the parameters and produces 4.5 answers/sec against TwIL-LM3's 28.1; the 120B is 40x the parameters and produces 3.4. TwIL-LM3 is the strongest arm here at its own scale and the most efficient arm at any scale.

The unreleased TwIL-LM3* moves the gate to 0.5896 and strict-7 to 0.3290, roughly +0.17 and +0.13 over the current release. The gains are concentrated in the two lanes where TwIL-LM3 is weakest in absolute terms rather than relative ones — rule_induction 0.3192 → 0.9644 and semantic_parse token-F1 0.4416 → 0.8762 — plus strict MCQ accuracy 0.1100 → 0.5200. It gives back lean_critic (0.6600 → 0.5200) and a little lm_corpus perplexity, so it is not uniformly better.

It is also the most efficient arm in the table by a wide margin — 28.1 answers/sec, from generations averaging 564 tokens where every other arm except Llama runs past 690. The Liquid models decode faster in raw tokens per second, 25230 and 22480 against 15880, but their length more than cancels it.

Track B — held-out benchmarks

dataset	TwIL-LM3	SmolLM3-3B base	Llama-3.2-3B	LFM2-2.6B	LFM2.5-8B-A1B	Qwen3-8B	gpt-oss-120b ‡
gsm8k	0.8733	0.8833	0.8300	0.8767	0.9133	0.9567	0.9767
svamp	0.8500	0.8567	0.8200	0.9000	0.9133	0.9400	0.9400
gsm_symbolic	0.7567	0.7633	0.8067	0.9767	0.9267	0.8133	0.8467
arc_cot	0.8467	0.8400	0.7967	0.8667	0.9033	0.9633	0.9667
logicbench	0.7167	0.6467	0.5733	0.6267	0.7200	0.8567	0.8533
strategyqa	0.6500	0.6333	0.6533	0.6433	0.6667	0.7400	0.7867
drop	0.7467	0.7000	0.6733	0.6900	0.6633	0.8833	0.8500
csqa	0.7367	0.7067	0.7500	0.7433	0.7700	0.8633	0.8367
musr	0.4957	0.4997	0.4932	0.4867	0.5703	0.6301	0.6852
mmlu_redux	0.6667	0.6633	0.6000	0.7133	0.8367	0.8500	0.9467
ifeval	0.6433	0.6767	0.7167	0.7300	0.8900	0.8400	0.7900
rudas_ood	0.0365	0.0209	0.0733	0.0017	0.0061	0.0468	0.0000 ¶
bbh_logic	0.6633	0.6667	0.5333	0.5713	0.7700	0.6367	0.9980
math500	0.6900	0.7000	0.4233	0.7133	0.7800	0.6100	0.8433
macro (10 CoT datasets)	0.7339	0.7193	0.6997	0.7523	0.7884	0.8493	0.8689
macro (all 14)	0.6694	0.6612	0.6245	0.6814	0.7378	0.7591	0.8086
tok/s	15880	15564	16160	25230	22480	9420	3374
mean gen length	482	626	510	≈796	≈1327	≈1931	801
ans/s	32.9	24.9	31.7	≈31.7	≈16.9	4.9	4.2

‡ MXFP4 weights, tensor-parallel 2 — quantized and multi-GPU, so not directly comparable to the single-GPU BF16 rows. ¶ 74% of its rudas_ood generations hit the length cap, so that cell is a truncation artefact rather than a measured score; excluding the row, its 13-dataset macro is 0.8708.

Lengths marked ≈ are derived from stored generations using each model's characters-per-token ratio rather than re-tokenized directly; the method reproduces the three directly measured lengths to within 3.5%.

The honest summary of this table is that TwIL-LM3 does not lead it. Larger models score higher, in order of size, and the 120B leads nine of fourteen rows. Two things are worth extracting anyway. First, TwIL-LM3 improves on its own base while sitting mid-table (0.7339 against 0.7193 on the 10-dataset macro), which is the point of the WiSE-FT stage — in-domain gains without transfer collapse. Second, it produces the shortest generations of any arm here at 482 tokens and consequently the most answers per second at 32.9, roughly eight times the 120B's rate.

Usage

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "webAI-Official/TwIL-LM3"
tok = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id, torch_dtype=torch.bfloat16, device_map="auto"
)

messages = [{"role": "user", "content":
             "Does 'All dogs are mammals. Rex is a dog.' entail 'Rex is a mammal'? "
             "Answer entailment, contradiction, or neutral."}]
inputs = tok.apply_chat_template(
    messages, add_generation_prompt=True,
    return_tensors="pt", return_dict=True,
).to(model.device)

out = model.generate(**inputs, max_new_tokens=2048, do_sample=False)
print(tok.decode(out[0][inputs["input_ids"].shape[-1]:], skip_special_tokens=True))

return_dict=True matters on transformers 5.x, where apply_chat_template returns a BatchEncoding rather than a bare tensor; the above works on both 4.x and 5.x.

The reported numbers use greedy decoding (do_sample=False) and a 2048-token generation budget. Note that the shipped generation_config.json inherits SmolLM3's sampling defaults (do_sample=true, temperature=0.6, top_p=0.95), so do_sample=False must be passed explicitly to reproduce the evaluation. The model opens a <think>...</think> reasoning block before answering, so a short generation budget truncates reasoning and scores far worse.

GGUF / llama.cpp

Quantized GGUF builds ship in this repository alongside the safetensors weights. The smollm3 architecture is supported by llama.cpp, and the chat template, <|im_end|> EOS and BOS are carried into the GGUF metadata, so chat mode works without extra flags.

file	quant	size	bits/weight	notes
TwIL-LM3-Q4_K_M.gguf	Q4_K_M	1.78 GiB	4.96	recommended default; runs on CPU or 4 GB of VRAM
TwIL-LM3-Q5_K_M.gguf	Q5_K_M	2.06 GiB	5.74	a little more headroom than Q4_K_M
TwIL-LM3-Q6_K.gguf	Q6_K	2.35 GiB	6.56	close to Q8_0 quality at two-thirds the size
TwIL-LM3-Q8_0.gguf	Q8_0	3.05 GiB	8.50	near-lossless, for quality-sensitive use
TwIL-LM3-F16.gguf	F16	5.73 GiB	16.00	unquantized, for requantization or reference runs

llama-cli -m TwIL-LM3-Q4_K_M.gguf -cnv --temp 0 -n 2048

Two things matter for reproducing the scores above under llama.cpp. Pass --temp 0, because the evaluation is greedy while the packaged sampling defaults are not. And leave the generation budget large — 2048 tokens or more — since the model emits a <think> block before answering and a short budget truncates it, which costs far more accuracy than the quantization does.

F16 and Q8_0 were produced directly by convert_hf_to_gguf.py from the released bf16 weights; the K-quants (Q4_K_M, Q5_K_M, Q6_K) were quantized from the F16 build with llama-quantize, without an importance matrix. All five were smoke-tested for load and generation on CPU. Note that F16 is not bit-identical to the released weights: bf16 and f16 carry the same 16 bits but trade exponent range against mantissa precision, so the conversion is a narrowing one, in practice negligible for inference.

The published Track A and Track B numbers were measured on the bf16 weights through vLLM, not on any of these GGUF builds, so expect small deviations — most likely at Q4_K_M — that have not been quantified here.

How it was built

Four stages on top of the base model:

1. LoRA supervised fine-tuning on a synthetic formal-logic corpus covering the Track A objectives (first-order-logic translation, entailment labelling, semantic parsing, Lean formalisation and critique, procedural reasoning, rule induction).
2. Checkpoint fusion — parameter-space averaging of intermediate SFT checkpoints selected by a diversity probe, rather than taking the final checkpoint.
3. WiSE-FT interpolation toward the pretrained base, W = (1 − λ)·W_base + λ·W_finetuned with λ = 0.25 — i.e. only a quarter of the fine-tuned delta is retained. λ was chosen by constrained optimisation: maximise in-domain score subject to minimal degradation on held-out benchmarks. This conservative λ is the direct reason held-out capability survives.
4. MGPO — entropy-weighted GRPO reinforcement learning against a programmatic verifier, with partial credit for loose matches and token-F1 so that all-fail prompt groups still produce gradient. Published checkpoint is step 2071.

A sibling arm that skipped stage 3's conservative interpolation scores considerably higher in-domain (macro gate 0.515) but gives back roughly twelve points of held-out capability. This release is the balanced point of that trade; the other was not published.

Limitations and caveats

Truncation. At a 2048-token budget, 4.4% of Track A generations hit the cap — better than the base's 17.4%, but still above the 2% threshold our protocol requires to mark a comparison rankable. The Track A macro gate should therefore be read as indicative rather than exact. Because a truncated response scores zero regardless of reasoning quality, both numbers are pessimistic, and the base substantially more so — meaning the true Track A gap is probably narrower than +0.086.

Scope. Tuned for formal logic. The Track B suite does not cover code generation or tool use (HumanEval, LiveCodeBench and BFCL were not run for this model or its base), so this release makes no claim about those.

Not a chat model. It was optimised against automatic verifiers on logic tasks. It has had no safety tuning beyond whatever the base model carries, and no instruction-following alignment work — IFEval regressed slightly.

Failed consolidation stage. A post-RL self-distillation round (SDFT) was attempted and made both tracks worse at every budget tried (−18% Track A at one epoch on this family). It is not part of this model. See the accompanying SDFT_RESULT.md in the project repository.

Evaluation protocol

• Track A: n = 200 per objective, greedy (temperature = 0), max_new_tokens = 2048, one retry at 4096 for truncated rows, max_seq_len = 8192, seed 42.
• Track B: 300 examples per task, greedy, max_gen_toks = 4096, max_model_len = 8192, repetition_penalty = 1.0, chat template applied, vLLM backend.
• Both tracks use the same protocol for the model and its base, in a paired run over identical sampled rows.

repetition_penalty = 1.0 is load-bearing. A 1.1 penalty produced apparent 20-point swings on Track B that were pure decoding artefact; the decoding kwargs are hashed into the protocol identity so a mismatched runner fails loudly instead of quietly producing a different number.

Relationship to TwIL-LM

webAI-Official/TwIL-LM is the 1.7B member of this family, built from SmolLM2 by the same pipeline. It reaches a higher in-domain score relative to its own base but gives back held-out capability; this model is the one that improves both. Unlike TwIL-LM's main branch, which ships a PEFT LoRA adapter, this repository ships a full merged model loaded directly with AutoModelForCausalLM.

License and attribution

Released under the webAI Non-Commercial License ver. 1.0 — see LICENSE.md in this repository.

The base model, HuggingFaceTB/SmolLM3-3B, is Apache 2.0; its licence text is retained as apache-2.0-LICENSE.txt and all credit for the base model goes to the HuggingFaceTB team. Apache 2.0 permits distributing derivative works under different terms provided attribution is preserved, which is what the pair of licence files in this repository does.