NOSIBLE/forward-looking-v1.2-base

Changelog

• v1.2.0: Multilingual upgrade. Extends coverage to 94 languages while holding English performance flat, and adds currency / G10-geography coverage.
• v1.1.0: English forward-looking model trained on real-world Nosible Search Feeds.

forward-looking-v1.2-base is a temporal-orientation classification model. Given a short text snippet, it determines whether the snippet's main event or topic is forward-looking (planned, expected, forecast, or scheduled) or not-forward-looking (a past event, a current state, or a timeless fact). It is fine-tuned from Qwen3-0.6B-Base and reframes the task as instruction following, producing a single label token per input.

This is the multilingual successor to forward-looking-v1.1-base. v1.1 was trained primarily on English; v1.2 extends the same task to 94 languages (English plus 93 additional languages) so the model can classify temporal orientation on text as it appears across global news and search feeds.

What's new in v1.2

• Multilingual coverage. The training corpus extends the English Forward-Looking data with faithful translations across 93 additional languages, where the forward / not-forward label is preserved through translation (future-tense framing is not flattened into the past, and vice versa).
• Wider topic coverage. v1.2 adds currency and G10-geography feeds, so temporal orientation is classified consistently across company, country / region, and currency news.
• English held flat. v1.2 is a multilingual extension, not an English re-train. English accuracy and macro-F1 are unchanged within run noise (see below).
• The multilingual gap narrowed. On the held-out validation set, the English-vs-multilingual accuracy gap shrinks from ~8.3pp (v1.1) to ~4.7pp (v1.2).

Performance overview

All numbers below are measured on the live SGLang endpoint (OpenAI-compatible chat-completions, enable_thinking=False, temperature=0), scored against the same held-out validation splits for both models. Deltas are in percentage points (pp).

Headline

Slice	n	Metric	v1.1	v1.2	Δ
English val	20,000	Accuracy	92.12%	91.96%	-0.16pp
English val	20,000	Macro-F1	91.91%	91.73%	-0.18pp
Multilingual val	19,155	Accuracy	83.83%	87.22%	+3.39pp
Multilingual val	19,155	Macro-F1	83.28%	86.92%	+3.64pp
Currency / geo feeds	4,012	Accuracy	86.86%	88.56%	+1.70pp
Currency / geo feeds	4,012	Macro-F1	86.18%	88.20%	+2.02pp

The small English regression (-0.16pp accuracy) is within per-run noise — v1.2 was a multilingual extension, not an English re-train, and the English baseline was already very strong. Multilingual accuracy improves by +3.39pp.

Selected languages (largest validation slices)

Language	n	v1.1 acc	v1.2 acc	Δ acc
Japanese (ja)	1,500	84.47%	87.80%	+3.33pp
Spanish (es)	1,271	87.96%	90.87%	+2.91pp
German (de)	1,603	87.40%	90.14%	+2.74pp
Italian (it)	669	87.44%	90.13%	+2.69pp
Russian (ru)	1,678	88.02%	90.58%	+2.56pp
French (fr)	1,225	89.06%	91.43%	+2.37pp
Portuguese (pt)	731	89.47%	91.52%	+2.05pp
Dutch (nl)	506	88.74%	90.12%	+1.38pp
Chinese (zh)	1,377	89.91%	90.92%	+1.01pp
Polish (pl)	588	82.14%	87.93%	+5.79pp

The gains are largest on lower-resource languages — for example accuracy rises on Tamil (61.95% → 75.22%), Gujarati (74.00% → 83.00%), and Swahili (60.53% → 67.54%), with even larger macro-F1 improvements.

Strict Usage Requirements

1. Disable Thinking: You must set enable_thinking=False (or disable reasoning tokens).
2. Exact System Prompt: You must use the specific system prompt: "Classify whether it is forward looking or not forward looking."
3. Constrain Output: You must restrict generation to the two label tokens the model emits: forward and _forward. The token _forward denotes the not-forward class — map it back after decoding.
   • SGLang: Use regex="(forward|_forward)" in the API call.
   • vLLM: Use guided_choice=["forward", "_forward"] in the API call.
   • llama.cpp / GGUF: Apply a GBNF grammar or regex to force selection from the two tokens.

Deviating from these requirements will severely impact performance and reliability.

The two classes are forward and not-forward. Internally the model was trained so that the not-forward class is emitted as the single token _forward. Always map _forward → not-forward (and forward → forward) after reading the model output.

Quickstart (local GPU)

Since this model was trained as a Causal LM using specific chat templates, you must use apply_chat_template with the exact system prompt used during training.

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "NOSIBLE/forward-looking-v1.2-base"

tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    device_map="auto",
    trust_remote_code=True,
    torch_dtype=torch.bfloat16,
)

# Multilingual input is supported (94 languages).
text = "El banco central elevará las tasas de interés el próximo mes."

# 1. Structure the prompt exactly as used in training
messages = [
    {"role": "system", "content": "Classify whether it is forward looking or not forward looking."},
    {"role": "user", "content": text},
]

# 2. Apply chat template (thinking MUST be disabled)
prompt = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
    enable_thinking=False,
)

inputs = tokenizer([prompt], return_tensors="pt").to(model.device)

# 3. Generate the label token (only a single token is expected)
outputs = model.generate(**inputs, max_new_tokens=1)
raw = tokenizer.decode(outputs[0], skip_special_tokens=True).split("<|im_start|>assistant\n")[-1].strip()

# Map the model-form token back to the human-readable label
label = "not-forward" if raw == "_forward" else "forward"
print(label)
# Expected Output: forward

Deployment

For production we recommend serving with SGLang (sglang>=0.4.6.post1), which exposes an OpenAI-compatible API endpoint. The model is based on Qwen3-0.6B and can be deployed anywhere Qwen3-0.6B can.

Launch the server:

python3 -m sglang.launch_server --model-path NOSIBLE/forward-looking-v1.2-base --dtype bfloat16 --host 0.0.0.0 --port 8080

Call the endpoint using the OpenAI-compatible client. Requesting logprobs lets you read a calibrated confidence for each label.

import math
from openai import OpenAI

# OpenAI-compatible client pointed at your SGLang server (set base_url to your
# endpoint URL if remote). The request shape mirrors signals_deploy_v12.predict_one.
client = OpenAI(base_url="http://localhost:8080/v1", api_key="EMPTY")

model_id = "NOSIBLE/forward-looking-v1.2-base"

# Multilingual input is supported.
text = "El banco central elevará las tasas de interés el próximo mes."

# The model emits the tokens `forward` / `_forward`; `_forward` == not-forward.
label_map = {"forward": "forward", "_forward": "not-forward"}

messages = [
    {"role": "system", "content": "Classify whether it is forward looking or not forward looking."},
    {"role": "user", "content": text},
]

completion = client.chat.completions.create(
    model=model_id,
    messages=messages,
    temperature=0,
    stream=False,
    logprobs=True,
    top_logprobs=2,
    extra_body={"chat_template_kwargs": {"enable_thinking": False}},  # Must be set to false.
)

# The top-1 token is the predicted label; map the model-form token back to the
# human-readable label. The full top_logprobs slice gives a per-label confidence.
top = completion.choices[0].logprobs.content[0].top_logprobs
print(f"Input: {text}")
print(f"Predicted Label: {label_map.get(top[0].token.strip(), top[0].token.strip())}")
print("--- Label Confidence ---")
for lp in top:
    name = label_map.get(lp.token.strip(), lp.token.strip())
    print(f"Token: {name!r} | Probability: {math.exp(lp.logprob):.2%}")

Expected Output

Input: El banco central elevará las tasas de interés el próximo mes.
Predicted Label: forward

--- Label Confidence ---
Token: 'forward' | Probability: 99.99%
Token: 'not-forward' | Probability: 0.01%

Legal Notice: This model is a modification of the Qwen3-0.6B model. In compliance with the Apache 2.0 license, we retain all original copyright notices and provide this modification under the same license terms.

Limitations

• Parameter Size (0.6B): As a small language model, it is designed for fast, specific classification and may struggle with highly nuanced or ambiguous text that requires extensive world knowledge.
• Per-language quality varies. Accuracy on the highest-resource languages approaches the English baseline; lower-resource languages remain below it despite the large v1.2 improvements.
• Domain Specificity: The model is fine-tuned on financial contexts and classifies temporal orientation, not general-purpose tense detection.
• Reporting-verb ambiguity: Temporal orientation is judged by the main event, not the tense of reporting verbs ("said", "announced"). Very terse or context-free snippets can be ambiguous.

Disclaimer

• Not Financial Advice: The outputs of this model should not be interpreted as financial advice, investment recommendations, or an endorsement of any financial instrument or asset.
• Risk: Financial markets are inherently volatile and risky. Never make investment decisions based solely on the output of an AI model. Always consult with a qualified financial professional.

Team & Credits

This model was developed and maintained by the following team:

• Matthew Dicks
• Gareth Warburton
• Stuart Reid

Citation

If you use this model, please cite it as follows:

@misc{nosible2025forwardlookingv12,
  author = {NOSIBLE},
  title = {Forward Looking v1.2 Base},
  year = {2025},
  publisher = {Hugging Face},
  journal = {Hugging Face Repository},
  howpublished = {https://huggingface.co/NOSIBLE/forward-looking-v1.2-base}
}

Full language breakdown

v1.1 was trained on English only; v1.2 adds the 93 languages below (94 total with English). The figures are the training-time evaluation per language and reproduce on the served SGLang endpoint to within ~0.2pp. Deltas are in percentage points (pp), sorted by validation row count.

Language	n	v1.1 acc	v1.2 acc	Δ acc	v1.1 F1	v1.2 F1	Δ F1
Russian (ru)	1,677	88.19%	90.64%	+2.45pp	88.00%	90.56%	+2.56pp
German (de)	1,603	87.34%	90.27%	+2.93pp	87.01%	90.12%	+3.11pp
Japanese (ja)	1,500	84.47%	87.93%	+3.46pp	83.92%	87.55%	+3.63pp
Chinese (zh)	1,377	89.83%	90.56%	+0.73pp	89.68%	90.47%	+0.79pp
Spanish (es)	1,271	87.96%	90.72%	+2.76pp	87.75%	90.55%	+2.80pp
French (fr)	1,225	88.82%	91.59%	+2.77pp	88.63%	91.46%	+2.83pp
Portuguese (pt)	731	89.74%	91.52%	+1.78pp	89.57%	91.39%	+1.82pp
Italian (it)	669	87.59%	89.69%	+2.10pp	87.40%	89.55%	+2.15pp
Polish (pl)	588	82.31%	88.27%	+5.96pp	81.83%	88.13%	+6.30pp
Dutch (nl)	506	88.74%	90.32%	+1.58pp	88.31%	89.99%	+1.68pp
Turkish (tr)	393	82.19%	86.77%	+4.58pp	81.75%	86.54%	+4.79pp
Indonesian (id)	386	87.56%	87.82%	+0.26pp	86.92%	87.32%	+0.40pp
Vietnamese (vi)	368	85.33%	89.95%	+4.62pp	85.20%	89.91%	+4.71pp
Czech (cs)	332	84.94%	85.84%	+0.90pp	84.72%	85.60%	+0.88pp
Korean (ko)	291	87.97%	88.32%	+0.35pp	87.95%	88.30%	+0.35pp
Arabic (ar)	289	87.54%	92.39%	+4.85pp	87.42%	92.36%	+4.94pp
Ukrainian (uk)	246	85.37%	91.46%	+6.09pp	84.79%	91.29%	+6.50pp
Swedish (sv)	228	83.77%	85.96%	+2.19pp	83.38%	85.78%	+2.40pp
Romanian (ro)	223	83.41%	85.65%	+2.24pp	83.18%	85.47%	+2.29pp
Hindi (hi)	191	83.25%	85.34%	+2.09pp	83.00%	85.22%	+2.22pp
Greek (el)	187	84.49%	87.17%	+2.68pp	83.58%	86.52%	+2.94pp
Thai (th)	179	88.83%	88.27%	-0.56pp	88.48%	87.97%	-0.51pp
Hungarian (hu)	177	76.27%	83.05%	+6.78pp	75.52%	82.76%	+7.24pp
Danish (da)	176	81.82%	88.64%	+6.82pp	81.35%	88.46%	+7.11pp
Malay (ms)	143	85.31%	88.11%	+2.80pp	84.73%	87.82%	+3.09pp
Slovak (sk)	143	83.92%	86.01%	+2.09pp	83.53%	85.76%	+2.23pp
Bengali (bn)	142	85.21%	88.73%	+3.52pp	84.57%	88.50%	+3.93pp
Persian (fa)	136	78.68%	80.15%	+1.47pp	77.16%	80.02%	+2.86pp
Finnish (fi)	136	78.68%	85.29%	+6.61pp	76.31%	84.43%	+8.12pp
Urdu (ur)	120	82.50%	87.50%	+5.00pp	81.83%	87.18%	+5.35pp
Norwegian (no)	116	89.66%	90.52%	+0.86pp	89.19%	90.26%	+1.07pp
Swahili (sw)	114	59.65%	65.79%	+6.14pp	44.19%	63.06%	+18.87pp
Serbian (sr)	113	86.73%	91.15%	+4.42pp	86.20%	90.99%	+4.79pp
Tamil (ta)	113	62.83%	75.22%	+12.39pp	61.97%	74.98%	+13.01pp
Hebrew (he)	110	80.91%	87.27%	+6.36pp	80.55%	87.23%	+6.68pp
Marathi (mr)	109	72.48%	75.23%	+2.75pp	71.61%	74.93%	+3.32pp
Bulgarian (bg)	107	85.05%	87.85%	+2.80pp	85.01%	87.78%	+2.77pp
Punjabi (pa)	105	74.29%	78.10%	+3.81pp	73.82%	77.57%	+3.75pp
Telugu (te)	102	62.75%	68.63%	+5.88pp	58.09%	67.10%	+9.01pp
Hausa (ha)	101	52.48%	60.40%	+7.92pp	47.53%	58.61%	+11.08pp
Gujarati (gu)	100	73.00%	83.00%	+10.00pp	70.88%	82.61%	+11.73pp
Tagalog (tl)	99	71.72%	71.72%	+0.00pp	67.13%	67.68%	+0.55pp
Kannada (kn)	93	64.52%	68.82%	+4.30pp	59.75%	64.01%	+4.26pp
Croatian (hr)	91	89.01%	91.21%	+2.20pp	88.90%	91.03%	+2.13pp
Azerbaijani (az)	87	85.06%	83.91%	-1.15pp	84.39%	82.70%	-1.69pp
Pashto (ps)	87	51.72%	71.26%	+19.54pp	37.94%	70.50%	+32.56pp
Uzbek (uz)	84	59.52%	72.62%	+13.10pp	55.24%	72.52%	+17.28pp
Burmese (my)	83	55.42%	81.93%	+26.51pp	47.64%	81.76%	+34.12pp
Nepali (ne)	83	74.70%	85.54%	+10.84pp	74.33%	85.52%	+11.19pp
Malayalam (ml)	82	75.61%	80.49%	+4.88pp	68.94%	77.91%	+8.97pp
Odia (or)	81	75.31%	79.01%	+3.70pp	72.71%	76.60%	+3.89pp
Kazakh (kk)	77	71.43%	77.92%	+6.49pp	69.58%	76.32%	+6.74pp
Somali (so)	77	70.13%	72.73%	+2.60pp	45.09%	65.10%	+20.01pp
Amharic (am)	75	53.33%	56.00%	+2.67pp	37.22%	42.83%	+5.61pp
Lithuanian (lt)	73	73.97%	78.08%	+4.11pp	72.66%	77.88%	+5.22pp
Sindhi (sd)	73	78.08%	83.56%	+5.48pp	76.10%	82.08%	+5.98pp
Sinhala (si)	68	66.18%	63.24%	-2.94pp	59.85%	51.91%	-7.94pp
Afrikaans (af)	66	87.88%	83.33%	-4.55pp	87.59%	83.14%	-4.45pp
Khmer (km)	66	40.91%	71.21%	+30.30pp	31.01%	70.88%	+39.87pp
Slovenian (sl)	66	78.79%	95.45%	+16.66pp	76.64%	95.32%	+18.68pp
Assamese (as)	63	76.19%	79.37%	+3.18pp	74.55%	77.05%	+2.50pp
Armenian (hy)	56	83.93%	82.14%	-1.79pp	82.30%	78.88%	-3.42pp
Kyrgyz (ky)	49	63.27%	75.51%	+12.24pp	56.76%	72.16%	+15.40pp
Latvian (lv)	48	81.25%	87.50%	+6.25pp	80.21%	87.30%	+7.09pp
Mongolian (mn)	46	65.22%	65.22%	+0.00pp	54.90%	54.90%	+0.00pp
Lao (lo)	44	81.82%	90.91%	+9.09pp	75.76%	87.88%	+12.12pp
Georgian (ka)	42	57.14%	69.05%	+11.91pp	51.79%	67.56%	+15.77pp
Sanskrit (sa)	22	72.73%	68.18%	-4.55pp	70.54%	66.45%	-4.09pp
Bosnian (bs)	21	85.71%	80.95%	-4.76pp	78.79%	76.67%	-2.12pp
Catalan (ca)	21	95.24%	90.48%	-4.76pp	95.06%	90.28%	-4.78pp
Irish (ga)	20	65.00%	65.00%	+0.00pp	49.82%	49.82%	+0.00pp
Belarusian (be)	19	73.68%	68.42%	-5.26pp	63.60%	59.29%	-4.31pp
Malagasy (mg)	19	57.89%	68.42%	+10.53pp	51.28%	66.07%	+14.79pp
Breton (br)	18	55.56%	61.11%	+5.55pp	44.62%	57.86%	+13.24pp
Welsh (cy)	18	44.44%	55.56%	+11.12pp	37.50%	55.00%	+17.50pp
Basque (eu)	18	77.78%	94.44%	+16.66pp	60.00%	92.59%	+32.59pp
Latin (la)	18	88.89%	88.89%	+0.00pp	71.88%	71.88%	+0.00pp
Macedonian (mk)	18	83.33%	83.33%	+0.00pp	83.28%	83.28%	+0.00pp
Oromo (om)	18	55.56%	55.56%	+0.00pp	44.62%	44.62%	+0.00pp
Serbo-Croatian (sh)	18	88.89%	94.44%	+5.55pp	87.50%	93.45%	+5.95pp
Xhosa (xh)	18	55.56%	50.00%	-5.56pp	35.71%	41.09%	+5.38pp
Yiddish (yi)	18	72.22%	83.33%	+11.11pp	41.94%	73.40%	+31.46pp
Galician (gl)	17	94.12%	100.00%	+5.88pp	94.12%	100.00%	+5.88pp
Icelandic (is)	17	47.06%	58.82%	+11.76pp	39.53%	47.11%	+7.58pp
Estonian (et)	16	68.75%	81.25%	+12.50pp	67.61%	81.18%	+13.57pp
Western Frisian (fy)	16	68.75%	68.75%	+0.00pp	67.61%	67.61%	+0.00pp
Scottish Gaelic (gd)	16	56.25%	56.25%	+0.00pp	45.89%	45.89%	+0.00pp
Javanese (jv)	16	75.00%	81.25%	+6.25pp	74.60%	80.57%	+5.97pp
Kurdish (ku)	16	68.75%	75.00%	+6.25pp	54.29%	66.67%	+12.38pp
Albanian (sq)	16	75.00%	75.00%	+0.00pp	74.60%	73.33%	-1.27pp
Sundanese (su)	16	81.25%	87.50%	+6.25pp	79.22%	87.30%	+8.08pp
Uyghur (ug)	16	68.75%	62.50%	-6.25pp	54.29%	38.46%	-15.83pp
Esperanto (eo)	15	80.00%	80.00%	+0.00pp	72.05%	76.19%	+4.14pp